A research programme (proposal and log on the project page) on recalibrating a camera from the features it already sees, with no target, and on whether the solver's own uncertainty is honest enough for a robot to accept or reject the update.
The free corners leave their best plane by 0.0077 square widths rms (0.024 at most), and the two cameras, solved separately, reconstruct the same bend (correlation 0.90; 0.91 with a second distortion term): the printed board is not flat, which a flat-board model cannot absorb (target-free reprojection error 0.26 against 0.30 px per coordinate). Over 400 random subsets of 3 to 12 views the focal length's z-score against the all-view reference has rms 1.94 (95% 1.75–2.12) for 4–12 views, so the reported σ is about twice too small; Zhang's is no better (2.16).
Limits: one rig and board, corners without mismatches, the reference is itself a flat-board fit, and every solve stopped on its evaluation cap.
Show the rest of the summary (49 more paragraphs, 4,245 words)
A pre-registered accept gate, σf/f ≤ 0.5%, rejected 92% of updates more than 2% off but accepted only 84% of those within 1%, short of the 90/90 target; an inflation factor fitted on one camera did not transfer cleanly to the other.
Milestone 3 asks whether the 2× is real and whether a robust σ fixes it. Scored against the target-free all-view solution with the nested variance, rms z is still 2.28, so the checkerboard reference is not the cause. A leave-one-view-out jackknife (3,480 re-solves) gives an honest σ with no fitted factor: rms z 0.70 and 0.65, 2σ coverage 97–99%, and 1.01 on near-critical subsets where the reported σ gives 2.34; the cluster-robust sandwich helps only on one camera.
But the jackknife is very conservative at 4 views (15× the reported σ) and ranks good against bad updates worse (ROC AUC 0.75 against 0.91), so a 0.5% jackknife gate accepts only 19% of good updates: an honest error bar and a good accept statistic are different things.
Milestone 5 tests the two suspects milestone 1 named. OpenCV's released-object calibration (a bend-aware Zhang) turns out to be the same estimator as the target-free bundle adjustment: identical intrinsics, σ and board shape to 1e-6 px, because free corners leave the board only its scale. Against it, the flat, evenly spaced board assumption alone moves the principal point cy by 8.7 px (left) and 5.3 px (right) while reporting a 1 px σ. A second radial term k2 is clearly non-zero but does not make σ honest: nested rms z 3.01 (2.00 without five outliers) against 2.28 with k1 alone. Residuals of neighbouring corners stay correlated (0.31 and 0.24), but only along one board axis (0.44–0.56, against about 0 along the other), which points at the board or the detector rather than the lens.
Milestone 1's reprojection comparison was corrected here (0.30 against 0.26 px per coordinate, not 0.42 against 0.26).
Milestone 6 traces the one-axis residual pattern to the corner localiser used so far: cornerSubPix with an 11×11 half-window (a 23 px box) misplaces corners on the board's outer ring by up to 6.4 px, because the box runs off the pattern. All 26 corners that differ by more than 1 px from a 5×5 window are on that ring. With a 5×5 window or OpenCV's findChessboardCornersSB, reprojection error halves and the one-directional structure mostly goes.
This corrects milestone 5: the flat-board shift of cy is 1.2 px (left) and 1.4 px (right) with SB corners, not 8.7 and 5.3; the bend survives, 25–45% shallower and still seen identically by both cameras. The localiser alone moves Zhang's focal length by up to 7.2 px (1.3%), several times any reported σ. And cleaner corners make the reported σ less honest, not more: it halves while the error barely moves, so the nested rms z rises from 2.28 to 3.1–3.5. A sandwich σ clustered by board column does not help (2.44 / 1.72). The four pre-registered hypotheses: one held, three refuted.
Milestone 4 turns this into the decision a robot makes. Each camera's 13 views are split into an old set of 7 and a new set of 3–6 other views (60 splits per size and camera, 908 solves), and a χ2 test on (f, cx, cy) decides whether to keep the old calibration, adopt the new one or flag for a target recalibration. On unchanged cameras the test with the solver's covariance fires on 72% of splits (nominal 1%): between disjoint view sets the covariance is 2.3–2.9× too small for f and cy and 4.6× for the right camera's cx, with no fall as views are added.
Inflated by the factor measured on the other camera, false alarms drop to 6.5%, but a 1% focal change (an exact model transform of the real corners) is detected in only 10–12% of splits with 5–6 views, and a real swap to the other camera unit (0.67% in f, 17 px in cx) in 30–77%. Even a factor fitted on the scored camera leaves about 4% as the smallest focal change detected most of the time. The rule adopts a harmful update in 1.4% of decisions with the transferred factor and 5.6% without. One of five pre-registered hypotheses held; the splits reuse 13 views, so the intervals are too narrow.
Milestone 4b replaces the borrowed factor with a covariance each calibration computes from its own views (8,580 more solves, two corner detectors). A leave-one-view-out jackknife makes the test quiet with nothing taken from another camera: 2.5% false alarms on the old corners and 1.7% on findChessboardCornersSB corners, against 58–74% for the solver's covariance and 36–38% for a sandwich clustered by view. It pays in power: with 6 new views it detects a 2% focal change in 7–12% of splits and 4% in 41–62%.
The detector matters as much as the statistics: with SB corners the jackknife test catches the real swap to the other camera unit in 75% of splits, against 30% with the old corners, and harmful adoptions are 0.5–1.5% of decisions. All five pre-registered hypotheses held.
Milestone 4c re-tests those stored solves with other thresholds (no new solve). A small-sample Hotelling reference (median 5.9–6.2 degrees of freedom) silences the test completely and keeps only a fifth of its 4% detections; even an oracle threshold fitted on the scored camera detects a 2% focal change in only 5–10% of splits with the three-parameter statistic, so there the limit is information. On the focal length alone the jackknife overstates the variance about 3× (99th percentile of z2 1.7–4.2, not 6.6): against the other camera's 99th percentile it detects a 2% change in 28–41% of splits and 4% in 78–85%, with 0–5% false alarms, close to the oracle, but it cannot see the unit swap, which lives in the principal point. Four of five pre-registered hypotheses held; the f-only power was not pre-registered, and the thresholds rest on 180 statistics from splits that reuse 13 views.
Milestone 6b settles which corner localiser is right by rendering each real view again from a known camera at its real pose, with blur and noise calibrated so the renders’ measured edge width (0.95–0.96 px) and noise match the real images’. On a board with full squares all three localisers find the corners to 0.04–0.06 px and the focal length to 0.04%, so four of five pre-registered predictions failed: the real 1% shift is not localiser behaviour on an ideal pattern.
The cause, found afterwards: the printed board’s end squares in two of its columns reach only 0.49–0.53 of a corner spacing past the last corner, against 0.94–0.97 elsewhere. Rendered that way, the 11×11 window is pulled up to 9.6 px towards the board’s edge and biases Zhang’s f by +0.4% to +1.5% (real images, against the 5×5 window: +0.64% / +0.91%), findChessboardCornersSB is biased the other way by −0.3% to −0.4% (real: −0.12% / −0.39%), and the 5×5 window stays within 0.04% of the truth. So Zhang on 5×5 corners (532.9 / 536.2 px) is the best focal length for this rig, and earlier milestones carry a localiser bias of +0.6–0.9% (11×11) or −0.1 to −0.4% (SB).
Limits: the renders leave out the bend, the clipboard and the sensor’s own noise, and the end-square measurement and ablation were not pre-registered.
Milestone 4d reruns the update test on those 5×5 corners (4,290 solves): the focal-length spread between disjoint view sets falls from 0.86–1.02% to 0.59%, and the f-alone jackknife test with the other camera’s threshold detects a 2% focal change in 61% of splits (34–35% on the biased corners) and the real unit swap in 79% (three-parameter test), but the borrowed threshold now gives 11% false alarms on the left camera, because the jackknife overstates the spread 1.5× on one camera and 2.8× on the other.
Milestone 4e asks whether a camera can set that threshold from its own old views (3,168 more solves: every 3- and 4-view subset). Splitting each 7-view old set into 4 + 3 in all 35 ways and scaling the halves’ focal-length spread by 1/n gives a test with nothing borrowed that is quiet (0.6% / 0.0% false alarms) but deaf: a 2% focal change is caught in 0.8% of splits. The halves overstate the real old-to-new spread 2.9× and 5.2×, because below 5 views the focal length’s spread falls much faster than 1/n (per view 3.3–5.2% at 3 views, 2.2% at 4, 1.0–1.7% at 5–7, found after the prediction failed). A per-unit threshold needs halves of at least 5 views, about 10 views in the old calibration. Two of five pre-registered predictions held, both only because the test stays silent.
Milestone 4f solves every subset of 3 to 12 of the 13 views on the 5×5 corners (16,172 view sets): the jackknife’s overstatement of the focal length’s real spread falls from 7–8× at 4 views to 1.1–1.4× at 10, mostly because a few delete-one solves of a small set land near the critical set, while the solver’s own σ is 2–6× too small at every size.
Milestone 4g scores the update test on every disjoint split those solves allow (37,752 per camera, no new solve). The borrowed threshold’s 61% was flattered by a 99th percentile from 180 statistics: on every split it catches a 2% focal change in 43% of 6-view updates (left 56–61%, right 26–30%). A robust jackknife (1.4826 × the MAD of the delete-one focal lengths) is still not an honest error bar (the right camera’s predicted spread is twice the real one), but against a plain χ21 at 1% it catches the 2% change in 54% at 2.3% false alarms, the same rate as the borrowed threshold, evenly on both cameras and with nothing taken from another unit. One of four pre-registered predictions held (the power one); the robust test’s left-camera false alarms, 3.4%, sit just above the 3% bar.
Milestone 4h shows that χ21 is the wrong reference for that robust test: if each view adds an independent offset to f, its z2 has a 99th percentile of 18.45 at 7 + 6 views and χ21 would give 6.1% false alarms, so the 2.3% here comes from two errors cancelling. Against a reference derived from that model, which borrows nothing and can be computed for any set sizes, the test stays at 0.45% / 0.20% false alarms but catches a 2% change from 6 views in 18.6% of splits (all three pre-registered predictions held).
Milestone 4i takes the reference from the unit itself: linearising its 13-view solve gives every view subset’s first-order response of f to corner error (within 0.2% of 9,600 nonlinear solves at 4–7 views), and drawing the unit’s own residual patterns through it gives each split a threshold (median 10.9 / 12.5 at 7 + 6 views, not 18.45). The twin explains most of the jackknife’s small-set inflation from geometry; against it the robust test fires on 1.44% / 0.37% of unchanged updates and catches a 2% change from 6 views in 33.1% of splits, with nothing borrowed.
The right camera’s extra spread, which its twin does not show, keeps it below the pre-registered 0.5% floor (three of four predictions held). With this reference the plain jackknife does as well (34.7%), and one threshold per design nearly as well as one per split (30.8%).
Milestone 4j runs the same test, unchanged, on two more cameras from the TUM RGB-D benchmark (a Kinect and an Xtion, 13 views each, 11,442 solves). Self-calibration lands within 0.64% of TUM’s calibration on the Xtion (+1.85% on the Kinect, whose lens needs more than k1). The twin reference stays at 0.06–0.80% false alarms with 5–6 new views but over-alarms with 4 (3.3% and 4.7%), where the delete-one sets are 3-view; a 2% change from 6 views is caught in 10% of splits, because these units fix f less precisely; and the real 1.4% focal difference between the two units is caught in 6–8%, as often as a scaled change of that size (one of five predictions held).
Milestone 4k re-runs all 11,442 TUM solves from the fleet’s spec focal length to remove the far minima, and finds they are not a solver artefact: the warm start reproduces all 38 far 6-view solves, which on the Xtion fit as well as the full calibration (rms 0.071 against 0.075 px), so those view sets leave f nearly free. On the Kinect the spec start lands the 13-view solve in a worse local minimum (rms 0.48 against 0.15 px).
A basin check that drops solves more than 10% from the spec raises false alarms at 4 new views to 20% on the Xtion and leaves the power at 10.6% (all five predictions refuted): keep the lowest-cost of several starts and never clip f to a window.
Milestone 4l asks whether the unit’s twin can say in advance which updates are able to catch a 2% focal change: its predicted power ranks the real detections with AUC 0.66–0.75 on four cameras, and testing only the top quarter doubles the TUM cameras’ power (10.0% to 21.5%) at 0–2.3% false alarms, while overstating the absolute power 1.2–1.9× (three of four predictions held).
Milestone 4m holds the old calibration fixed, as a robot must, and lets the twin choose among the candidate new view sets: its pick catches a 2% focal change 1.33× as often as a blind pick on the TUM cameras with 7 candidates (7.3% to 9.7%) and 1.63× with 28, 1.17× on OpenCV’s, at 0.06–1.40% false alarms; the twin predicts that gain well though not the level, and some candidate would have caught the change in 19–55% of old sets (three of five predictions held).
Milestone 4n refits the TUM Kinect, whose lens k1 alone cannot follow, with k1 + k2 (9,440 new solves): a 2% focal change from 6 new views is caught in 19.3% of 7 + 6 splits against 9.8%, and a real unit swap in 16–17% against 6–8%, at 0.59% false alarms; on the undistorted Xtion the extra term changes nothing (one of four predictions held: the expected precision cost did not appear).
Milestone 4o re-runs the twin’s ranking and waiting rule with k1 + k2: on the Kinect the twin’s predicted power now matches the real rate (0.995×, from 1.34×), and the gains multiply, so fitting k2 and then testing the twin’s pick among 7 candidate new sets catches a 2% focal change in 17.9% of old sets against 6.2% blind with k1 (×2.86) at 0.59% false alarms; on the Xtion the twin still overstates its power 1.96× (all six predictions held).
Milestone 4p scores the test on views it has never seen (a disjoint 13 per TUM unit, 9,440 new solves): false alarms stay under the 2.5% bar (0.9% Kinect, 1.9% Xtion) but above the in-sample 0.6% and 0.1%, and a 2% focal change is caught in 8.0% of splits against 14.8% in-sample; the fresh views were flatter, and on them alone the Xtion over-alarms (4.4%), so the test’s safety depends on the views as well as the threshold (three of four predictions refuted).
Milestone 4q lets the robot abstain: the test runs only when the twin’s first-order sd of the focal difference is at most 1% of f. On the same pools, power for a 2% change rises from 9.1% to 15.9% of scored splits (×1.75) and no split holding a far solve is scored, but the flat Xtion pool still over-alarms (4.4% to 5.9%) and the Kinect’s flat pool is never scored, so the over-alarm is not the far solves the twin cannot represent but a per-view error the twin understates (one of four predictions held; a check on the pools that suggested it, not on unseen data, since the TUM host refused a third pool).
Milestone 4r adds a per-view error to the twin: each view gets its own effective focal length, with its sd fitted per pool from nothing but the scatter of that pool’s own 6-view solves (0.2–0.75% of f on tilted pools, 1.8–1.9% on flat ones). Every unit and pairing then stays under 0.6% false alarms (the flat Xtion pool from 4.4%), but a 2% focal change is caught in 4.1% of splits instead of 9.1% (×0.45); with milestone 4q’s gate the robot tests 15% of splits and catches 15.9% of 2% changes there at 0.3% false alarms.
It also withdraws a reading: the Xtion’s two 13-view calibrations differ by 1.6 of the twin’s own sd, not 2.7, which had used an iid sd (three of four predictions held; checked on the pools that suggested it). From natural-feature (SIFT) tracks of ordinary hand-held video instead of a board (milestone 3b, three TUM RGB-D sequences, 57 counted 1.2-second windows), a single window’s focal length is off by a median 2.8–7.2% and the solver’s sd is 4–12× too small (rms z 9.3, so H5 fails), and it does not grow on near-critical translation-only motion; the median over 15–24 windows is within 1.5% of the board calibration on every sequence.
Milestone 4s tests milestone 4r’s jitter on a third disjoint 13 views per TUM unit (9,440 new solves), which played no part in suggesting it: every pairing holding the Xtion’s third pool stays at or below 0.53% false alarms, but the Kinect’s third pool, well tilted and fitted with a small jitter (0.77% of f), over-alarms at 2.9–7.5% whatever jitter the twin is given (borrowed from the other two pools: 2.9–7.2%). Its 7-view solves hold 62 far solutions where the first pool’s hold one, and splits holding a far solve carry 83–97% of its false alarms (without them 0.49%, a diagnostic added after scoring): a robust spread cannot see that error, so a jitter fitted to it cannot fix it. The three 13-view calibrations of each unit agree within 1.4 of the jittered twin’s sd (two of five predictions held).
Milestone 4t: withholding the decision whenever the old or new small-set focal length lies more than 5% from the fleet’s spec-sheet value (527.1 px) brings every one of the 18 pairings of three pools on both cameras to at most 1.11% false alarms (pooled 0.30%, from 1.61%; the Kinect’s third pool from 7.47% to 0.32%) while withholding 5.8% of unchanged-camera splits, but it catches a ±2% change in 3.9% of splits against 5.3% (0.73×) and withholds most 4% changes on the side where a unit’s focal length already sits from the spec (three of five predictions held).
Milestone 4u: centring the same 5% window on the unit’s own 13-view calibration instead keeps all 18 pairings at or under 1.05% false alarms (pooled 0.26%, 5.4% withheld), withholds a 4% change 19–22% of the time on both sides of both cameras (the spec-sheet window: 11–71%), and catches ±4% changes in 14.2% of splits against 11.4%; ±2% power is unchanged at 0.73×, so that loss belongs to the far solves, not the window (five of six predictions held).
Milestone 3c: on the same natural-feature video windows, a delete-one-keyframe jackknife (12 warm-started re-solves per window) gives an sd of f that is honest overall (rms z 1.76 over 57 windows, from the solver’s 9.31; 2-sd coverage 89%) and matches the real across-window scatter in every sequence (0.9–1.6×), while a track-clustered sandwich does not (rms z 6.6) and residuals barely correlate along tracks, so the missing error enters per keyframe. The honest sd is a median 6.4% of f, only 9% of windows reach 2%, and on the near-critical translation-only sequence the jackknife is still too small (rms z 2.58) and grows only 1.54× (two of six predictions held).
Milestone 3d: pooling those windows by DerSimonian–Laird random effects keeps the error bar honest (rms z 1.10 over five-window blocks; treating windows as independent gives 2.79), and about eight windows (10 s) reach a 2.5% sd.
Milestone 3e: the camera’s optical-axis swing, read from the solved poses (Spearman 0.98 against motion capture), does not predict a window’s focal error (Spearman +0.08) and is smallest on a sequence that pools well, so a motion gate on it buys nothing (one of five predictions held).
Milestone 3f: a window’s k1 tracks its focal error (Spearman +0.59), but conditioning f on the board’s k1 alone raises the median error from 5.25% to 7.01% (+12.4% on the near-critical sequence), because windows and board agree on the distortion curve and differ only in how k1 and k2 share it; it helps only on the nearly undistorted camera (five pooled windows 2.36% to 1.43%), so a prior calibration should be pinned whole (two of four predictions held).
Milestone 3g: pinning k1 and k2 together (first order, from the full covariances the solves stored) is worse still: the median error rises to 8.72% (+12.9% on the near-critical sequence) and the conditional bar narrows by about 40%, so pooled windows become confidently wrong (rms z 3.07); 3f’s “pin it whole” rule does not survive, and distortion is best left free in short windows (one of five predictions held).
Milestone 3h: on the real stereo images, where the pinned bundle adjustment can be run (149 windows, 1043 re-solves), 3g’s first-order formula gets the direction right in 99% of windows and the size within a few percent for small steps, but overshoots larger ones (the re-solved step is 0.74 of the predicted one at 5–10% of f); the narrowed bar is real, and shrinking 3g’s steps by the overshoot still leaves the board prior worse than none (7.1% against 5.25%), so its warning stands (four of five predictions held).
Milestone 3i: is the principal point what a short window’s f compensates? No. Pinned at the board’s value, it lowers the focal error on clean board windows (re-solved, 0.55% to 0.37%, first order accurate) but raises it on the 57 TUM natural-feature windows (5.25% to 6.76%, first order), whose principal points wander up to 33 px from the board’s while f barely leans on them (correlation −0.1 to −0.3); pinning principal point and distortion together is 7.17%. Neither static intrinsic explains the short-window error (three of seven predictions held).
Milestone 3j re-solves those 57 TUM windows by bundle adjustment with the board’s values pinned, now that the TUM host answers: the distortion doubles the median focal error (5.25% to 10.5%, f above the board’s in 52 of 53 windows; first order had said 8.4%), the principal point is mildly worse (5.87%; it helps only the far fr3 scene), all four 10.7%; first order gets the step’s median size right (re-solved/predicted 0.87–1.01) but its direction wrong in a quarter of the distortion re-solves, against 1% on board windows (five of six predictions held).
Milestones 3l and 3m put the gyro a robot already carries into the window solve as a rolling-shutter term (synthetic, 30 hand-held windows at 0.3 rad/s, 1/30 s readout): the median focal shift falls from 4.17% to 1.14% and one window estimates the readout to 3.6%; the fix tolerates a 10° camera–IMU rotation (0.42% added) and a 10 ms time offset (0.35%), but a readout or gyro scale held fixed and 5% off adds 1.23%, in the same direction in all 30 windows, which a free readout absorbs exactly. Fault sizes are swept, not sourced.
Milestone 3n runs that fix on the real TUM video, with per-keyframe angular rates from the 100 Hz motion capture and the readout free: both cameras show a rolling shutter (Kinect 27.6 ms, Xtion 25.9 ms; positive in all 54 windows, about 0 with rates from the wrong time), and the median focal error falls from 2.87% to 1.88% (fr1/desk), 5.00% to 3.96% (fr1/xyz) and 7.56% to 3.63% (fr3); the Xtion’s readout depends on how the mocap rate is differenced (19–36 ms), and a free linear velocity did not converge.
Milestone 3o (synthetic) frees the gyro–camera time offset in the same window: a 20 ms late gyro’s focal shift falls from 0.81% to 0.05% and the readout stays unbiased, but the offset itself lands about 3 ms off (solver claims 0.46 ms) because it absorbs the unmodelled linear velocity; solving for the velocity brings it to 0.5 ms.
Milestone 3p (real TUM video) puts the motion-capture linear velocity into the window as a known input and frees the mocap–camera offset: the velocity fits the images slightly better (cost lower in 26 of 30 Kinect windows, beating another window’s velocity in 23 of 28) but does not move the focal error (paired Wilcoxon p = 0.44 over 53 windows); the offset is real, a median 5.1 ms with the camera’s stamps behind the motion (positive in 46 of 53 windows, sign test p = 4e-8; 6.0, 5.4 and 4.0 ms by sequence), and milestone 3o’s synthetic velocity bias does not appear on TUM (0.55 ms).
Milestone 3r (real TUM video) adds a soft prior on k1 and the principal point, set to the median and spread of the same camera’s other windows (no target, no reference value): the window-to-window focal scatter falls by a quarter to a half, the pooled median focal error from 3.06% to 2.43% (lower in 35 of 53 windows, paired Wilcoxon p = 0.002), windows over 10% wrong from 9 to 4, the median focal length moves under 1% and the image cost 0.05%; a prior from the camera’s other recording session works as well, k1 alone carries the gain, and the Xtion’s +2% bias stays.
Show 88 earlier entries
2026-11-19 · Milestone 3r planned: on TUM, does a prior on k<sub>1</sub> and the principal point drawn from the same camera's other windows narrow the window-to-window focal scatter that milestone 3q found, without moving the centre?
Written before experiments/m3r_prior/run.py was run on any TUM window and before analyse.py was written. Milestone 3q found the Kinect's leftover error is scatter (robust sd 2–6%) along an f–k<sub>1</sub>–principal-point ridge, and milestones 3f–3j found that pinning k<sub>1</sub> at the board value hurts. A robot with a fleet memory has a third option: a soft prior from its own earlier self-calibrations, which uses no target. Seen before this plan: the medians and robust sds of milestone 3p's est intrinsics over the 53 counted windows (Kinect c<sub>x</sub> 319.2 ± 9.8, c<sub>y</sub> 259.0 ± 7.5, k<sub>1</sub> 0.127 ± 0.040; Xtion 337.0 ± 20.9, 239.2 ± 24.3, −0.018 ± 0.050), which set the priors below; nothing from any prior solve.
- Method (
run.py). Milestone 3p's windows, nfba start and counting rule; milestone 3n's est model (mocap rates, u = 0, readout free from 0, all five intrinsics free) plus Gaussian prior residuals r<sub>j</sub> = σ<sub>px</sub>(p<sub>j</sub> − μ<sub>j</sub>)/s<sub>j</sub> on chosen intrinsics, σ<sub>px</sub> the window's own est rms (px), so the prior is in units of the window's pixel noise. μ and s are the median and robust sd (1.4826 MAD) of milestone 3p's est values over the other counted windows of the same camera (leave-one-window-out; no reference calibration enters). Models: est_p (c<sub>x</sub>, c<sub>y</sub>, k<sub>1</sub>, width s), est_ph (same, width s/2), est_pk1 (k<sub>1</sub> only, width s), est_px (Kinect only: c<sub>x</sub>, c<sub>y</sub>, k<sub>1</sub>, width s, μ and s from the other fr1 sequence: a prior from an earlier session). est itself is taken from milestone 3p; check: the nfba start's f equals milestone 3p's to 1e-6 relative. e = (f − R)/R against milestone 4n's board calibration (fr1 518.28, fr3 530.71 px) unless stated. - H208 (the prior narrows the scatter). Robust sd of e for est_p ≤ 0.75 × est's, in at least 2 of 3 sequences.
- H209 (and lowers the error, paired). Pooled over counted windows: median |e| of est_p ≤ 0.85 × est's, and a Wilcoxon signed-rank test of |e<sub>est_p</sub>| − |e<sub>est</sub>| gives p < 0.05 with |e| lower in more than half the windows.
- H210 (the centre does not move). Per sequence, |median f(est_p) − median f(est)| / median f(est) ≤ 1%, in all 3.
- H211 (a prior from an earlier session carries). est_px: median |e| ≤ 0.9 × est's on both fr1/desk and fr1/xyz.
- H212 (tighter helps). Pooled median |e|: est_ph ≤ est_p.
- H213 (the principal point carries weight, not just k<sub>1</sub>). Pooled median |e|: est_p ≤ 0.95 × est_pk1.
- H214 (the images accept the prior). Pooled median rise in the image (robust, prior excluded) cost of est_p over est ≤ 1%.
- Reported, unscored. All of the above against TUM's published and the target-free references; per-sequence medians and robust sds of f, k<sub>1</sub>, c<sub>x</sub>, c<sub>y</sub> by model; solver sd<sub>f</sub> by model against the observed scatter; the Xtion's median e with the prior;
prior.png. - Caveat set in advance. The leave-one-out prior comes from windows of the same sequences (same session, scene and temperature), so H208–H210 measure a best case; est_px is the honest cross-session test but exists only on the Kinect; the prior's width is the population scatter, which includes the scatter it is meant to remove; 53 windows, three sequences, two cameras.
2026-11-18 · Milestone 3q result: on TUM, what is left after the shutter term is scatter between windows, not a bias (robust sd 2.1–6.0% against a median 0.1–2.5%); the Kinect windows' median sits inside the three references' 1.6% disagreement; k<sub>1</sub> and the principal point track a window's error, the solver's sd barely does; four of seven predictions held
- Run.
experiments/m3q_residual/analyse.py (no new solves; seconds) on milestone 3p's 53 counted windows (summary.json, residual.png), as planned below. tests/test_m3q.py (3 tests) passed. - H201 held. Median f of est: fr1/desk 524.1 px (bootstrap 95% 518.3–531.5), fr1/xyz 517.9 (511.4–538.9), both inside the Kinect references' span 516.9 (TUM) to 525.2 (target-free); fr3/stf 544.2 (527.7–565.6), above the Xtion's span 530.7 (board) to 537.3 (TUM), though its interval covers all three.
- H202 held. Robust sd of e against |median e| (board): fr1/desk 2.1% against 1.1%, fr1/xyz 6.0% against 0.07%, fr3/stf 6.0% against 2.5%.
- H203 held. fr1 pooled Spearman of signed e with the window's k<sub>1</sub> +0.58 (30 windows; milestone 3f without the shutter term +0.59). Within each sequence too (unscored): +0.44, +0.60, and +0.72 on fr3, whose images come undistorted.
- H204 refuted, narrowly. Pooled Spearman of |e| with sd<sub>f</sub>/f +0.29 (bar +0.3); +0.21 against TUM's reference, +0.35 against the target-free one, +0.21 for est_mud.
- H205 refuted. Keeping the half of each sequence's windows with the smaller sd<sub>f</sub>/f: median |e| 2.11% → 1.65% (fr1/desk, ratio 0.78), 3.96% → 4.20% (fr1/xyz, 1.06), 3.63% → 2.31% (fr3, 0.64). One of three sequences reached the 25% bar.
- H206 held. Pooled Spearman of |e| with the window's principal-point distance from the target-free anchor +0.42 (median distance 20.6 px); +0.38 against TUM's and target-free references, +0.31 to +0.38 for est_mud.
- H207 refuted. A line e = a + b(k<sub>1</sub> − k<sub>1,board</sub>) fitted on fr1/desk and applied to fr1/xyz halves its median |e| (3.96% → 2.12%, ratio 0.54); fitted on fr1/xyz and applied to fr1/desk it makes it worse (2.11% → 3.51%, 1.66). The slopes differ (0.73 and 1.24 per unit k<sub>1</sub>) and so do the intercepts (7.2% and 12.1%), so the trade-off is real in each session but its line is not the camera's.
- Unscored. Pooled median |e| of est: 3.06% (board), 2.96% (TUM), 2.62% (target-free); for est_mud 2.79%, 3.02%, 2.05%. The choice of reference moves the headline figure by up to 0.7 pp, as much as any model change since milestone 3n.
- Reading. Milestones 3n–3p asked what bias the window's f still carries; this one finds there is little bias left to explain on the Kinect. Its windows' median sits inside the 1.6% that three honest calibrations of the same camera disagree by, and the error is the window-to-window spread (2–6% robust sd), which tracks the window's own k<sub>1</sub> (ρ +0.44 to +0.72 in every sequence) and principal point (+0.42): within one short window, focal length, distortion and principal point trade off along a ridge, and where a window lands on it is what sets its error. The Xtion's median is 1.3–2.5% high against every reference, the one bias that remains. For a practitioner: a single window's f is good to a few percent, and its own solver sd is a weak guide to which windows to trust; pool windows (milestone 3d) and quote the reference's own uncertainty, about 1–2% on TUM, beside any claim below that. A fixed k<sub>1</sub>–f correction learned in one session does not carry to the next.
- Caveat. 53 windows, three sequences, two cameras; the principal-point reference is the target-free anchor because the Zhang fit's principal point is not stored; H205's gate uses each sequence's own median; the paired correction in H207 is fitted on 13 or 17 windows.
2026-11-18 · Milestone 3q planned: on TUM, with the shutter term in, is the 1.9–4.0% focal error left per window a bias against the board reference or per-window scatter, and does anything the window itself reports (its k<sub>1</sub>, its solver sd, its principal point) predict it?
Written before experiments/m3q_residual/analyse.py was written or run. No new solves: the inputs are milestone 3n's est (shutter term in) and milestone 3p's est_mud (offset and velocity in) on milestone 3p's 53 counted windows. Seen before this plan: milestone 3n's and 3p's summaries as logged (per-sequence median |e| and median e, e.g. fr1/desk est median e +1.06%), and the three references below; no per-window value of k<sub>1</sub>, sd or principal point against e.
- References (all with the windows' k<sub>1</sub> + k<sub>2</sub> model unless noted): R<sub>board</sub> = milestone 4n's flat-board Zhang, fr1 518.28, fr3 530.71 px (the reference of every earlier milestone); R<sub>TUM</sub> = the dataset authors' published calibration, (f<sub>x</sub> + f<sub>y</sub>)/2, fr1 516.9, fr3 537.3 (Sturm et al. 2012); R<sub>tf</sub> = milestone 4n's target-free 13-view anchor, fr1 525.16, fr3 534.04. e = (f − R)/R; R<sub>board</sub> unless stated.
- H201 (the window median sits inside the references' disagreement). For each sequence, the median f of est lies between the smallest and largest of the three references, in at least 2 of 3 sequences.
- H202 (what remains is scatter, not bias). For each sequence, the robust sd (1.4826 MAD) of est's e across windows exceeds |median e|, in all 3 sequences.
- H203 (milestone 3f's k<sub>1</sub>–f trade-off survives the shutter term). fr1 pooled (desk + xyz): Spearman of signed e (est) with the window's k<sub>1</sub> ≥ +0.4 (milestone 3f, no shutter term: +0.59).
- H204 (the solver's own sd flags bad windows). Pooled over 53 windows: Spearman of |e| with sd<sub>f</sub>/f (est) ≥ +0.3.
- H205 (a gate a robot could run). Keeping only windows with sd<sub>f</sub>/f at or below their sequence's median lowers median |e| by ≥ 25% against all of that sequence's windows, in at least 2 of 3 sequences.
- H206 (principal point). Pooled: Spearman of |e| with the window's distance of (c<sub>x</sub>, c<sub>y</sub>) from milestone 4n's target-free k<sub>1</sub>+k<sub>2</sub> anchor's (fr1 313.3, 262.5; fr3 331.2, 251.7; the Zhang fit's are not stored) ≥ +0.3.
- H207 (a k<sub>1</sub> correction transfers between sessions of one camera). On fr1, fit e = a + b(k<sub>1</sub> − k<sub>1,board</sub>) by least squares on one sequence and apply it to the other (desk → xyz and xyz → desk); median |e| corrected ≤ 0.8 × uncorrected on both held-out sequences.
- Reported, unscored. Every figure above against R<sub>TUM</sub> and R<sub>tf</sub> too, and for est_mud; per-sequence medians of f with a bootstrap 95% interval (2,000 resamples, seed 0);
residual.png. - Caveat set in advance. 53 windows from three sequences of two cameras; windows of one sequence share a scene and session; H205's gate uses the sequence's own median, which a robot has only after it has seen the windows; the references are themselves estimates whose spread (fr1 1.6%, fr3 1.2%) is the point of H201.
2026-11-18 · Milestone 3p result: on TUM, the motion-capture velocity fits the images slightly better but does not move the focal error; the mocap–camera offset is real (a median 5.1 ms, positive in 46 of 53 windows) and does not need the velocity to find it; three of six predictions held
- Run.
experiments/m3p_tumvel/run.py on milestone 3n's 64 windows, about 60 min on 4 cores, then analyse.py (summary.json, tumvel.png), as planned below. 53 windows counted: fr1/desk 13, fr1/xyz 17, fr3/stf 23 (milestone 3n's 54 less fr1/desk window 16, whose rate signal, padded 0.1 s past the window, runs into a mocap gap). Every counted window's est, est_mu, est_d and est_mud solve converged; mu_shuf is missing in 3 windows whose source window has a gap. tests/test_m3p.py (3 tests) passed before the run. Median scale 4.7 window units per metre (Kinect) and 4.9 (Xtion); median mocap spread of the keyframe centres 0.12 m and 0.07 m; median speed 0.37 and 0.21 m/s. - Check held. est's f equals milestone 3n's est exactly in all 53 windows (median and max relative difference 0).
- H195 refuted, narrowly. est_mu's cost is below est's in 26 of 30 Kinect windows (87%) but 17 of 23 Xtion windows (74%; bar 75%). The drops are small: a median 0.8% (Kinect) and 0.2% (Xtion) of the robust cost.
- H196 held. est_mu's cost is below mu_shuf's in 23 of 28 Kinect windows (82%) and 17 of 22 Xtion windows (77%). The velocity's timing, not just its size, is what fits.
- H197 held by its letter, not supported by the paired test. Kinect pooled median |e| 2.14% with the velocity against 2.62% without (ratio 0.82, bar 0.9). But |e| falls in only 15 of 30 windows, the median paired change is −0.09 pp (a slight rise), and a Wilcoxon signed-rank test gives p = 0.45 (all 53 windows: p = 0.44, median change −0.05 pp). The medians moved because a few windows on either side of the median swapped places (fr1/xyz: 3.96% → 2.21% by median, |e| lower in 9 of 17). Read: the mocap velocity does not change a window's focal error by a detectable amount. (The paired tests were added to
analyse.py after scoring.) - H198 held. est_mud's offset: fr1/desk median 6.1 ms, robust sd 3.5 ms; fr1/xyz 5.4 ms, 4.9 ms; fr3/stf 4.0 ms, 2.9 ms (bars 10 ms and 5 ms). Unscored: positive in 46 of 53 windows (sign test p = 4e-8); pooled median 5.1 ms, robust sd 4.0 ms. The solver claims a median sd of 0.7 ms (Kinect) and 0.2 ms (Xtion), so the window-to-window scatter is again 5–15× what it reports. With δ̂ > 0 the camera's motion at its stamp is what the mocap reports about 5 ms earlier: TUM's colour-image stamps run about 5 ms behind the instant the image shows, on both cameras.
- H199 refuted. Kinect median |δ̂<sub>est_d</sub> − δ̂<sub>est_mud</sub>| 0.55 ms (bar ≥ 2 ms; fr3 0.28 ms). Milestone 3o's synthetic 3 ms bias from leaving the velocity out does not appear on TUM. Its synthetic windows (milestone 3k's part B) put more translation into the smear than these hand-held and slow sequences do, so on TUM the rotation carries the offset's signal; the offset scatter is not from the missing velocity.
- H200 refuted. Freeing the offset moves f a median 0.45% with the velocity in (Kinect; fr3 0.58%; bar 0.3%). Not consistently in one direction: est_mud's |e| is below est's in 30 of 53 windows (57%); median |e| fr1/desk 2.11% → 1.62%, fr1/xyz 3.96% → 3.45%, fr3 3.63% → 3.14%.
- Unscored. Readouts barely move: Kinect 27.4 ms (est) and 25.3 ms (est_mu), Xtion 25.9 and 26.2 ms; the velocity tightens the Kinect's window-to-window readout spread from 3.9 to 2.8 ms and the Xtion's from 5.2 to 3.7 ms. est_d (offset free, no velocity) also lowers the cost against est in 80% (Kinect) and 96% (Xtion) of windows. Spearman of the per-window gain in |e| from the velocity against speed: −0.35 (fr1/desk), +0.13 (fr1/xyz), +0.01 (fr3).
- Reading. Milestone 3n left 1.9–4.0% of median focal error after the shutter term. This milestone tested two of the remaining modelled causes on the real data and neither is it: the camera's translation during readout is visible in the images (the true velocity fits better than another window's) but is too small at these speeds to move f, and a ~5 ms timing offset is there but freeing it moves f by about half a percent in either direction, not systematically toward the board. What remains is not a shutter-timing effect; the leads are the per-window geometry (milestone 3b–3j: the spread of view angles predicted the error) and the board reference itself. For a practitioner on TUM-class RGB-D data: the mocap–image offset is about +5 ms and is worth entering; the mocap velocity can be left out of a short self-calibration window; and, like the readout, the offset a single window reports is good to about 3–5 ms, not to the solver's sub-millisecond sd.
- Caveat. The velocity's scale comes from each window's own nfba start (a short baseline gives a noisy s); the velocity is a 40 ms difference held at the camera stamp; the offset is relative to TUM's mocap stamps after their own synchronisation, not a hardware latency; three sequences, two sensor units; H197's median test was weaker than the paired test that should have been set in advance.
2026-11-18 · Milestone 3p planned: on TUM, does the motion-capture linear velocity, entered as a known input, explain more of a 1.2 s window's images and its focal error, and does a free mocap–camera offset stay put once it is in?
Written before experiments/m3p_tumvel/run.py was run on any TUM window and before analyse.py was written. Milestone 3n's free-velocity model converged in 6 of 54 windows, and milestone 3o found on synthetic windows that a free time offset with no velocity in the model is biased by about 3 ms (0.5 ms once the velocity is in). TUM's ground truth gives the optical centre's position, so the velocity can be an input rather than 36 unknowns. Seen before this plan: tests/test_m3p.py (3 tests: the velocity's sign against a moving, turning camera; the scale; a mocap gap) passed; nothing from any TUM solve of this milestone.
- Method (
run.py). Milestone 3n's windows, nfba start and counting rule (converged, ≥ 100 tracks, inlier rms < 1.5 px, 300 < f < 1000, defined sd, mocap at every keyframe and over the rate signal). From the nfba solution, all five intrinsics and the readout free from 0, five solves: est (milestone 3n's model, u = 0), est_mu (u<sub>i</sub> = −s R<sub>i</sub><sup>T</sup>c′(t<sub>i</sub>) from mocap, c′ a ±20 ms difference; s, window units per metre, the ratio of the rms spread of the nfba camera centres to that of the mocap centres), mu_shuf (the velocity of the window five on, same scale s: a null with the velocity's size but not its timing), est_d (milestone 3o's offset model on a 5 ms mocap rate signal, u = 0, δ̂ free from 0), est_mud (both). δ̂ > 0 means the camera's motion at its stamp is what the mocap reports δ̂ earlier. e = (f − f<sub>ref</sub>)/f<sub>ref</sub> against milestone 4n's board calibration, as milestone 3n. - Counted. As above; a solve that does not converge (status 0) is reported and left out of that model's figures; paired comparisons use windows where both solves converged.
- Check, not a hypothesis. est's f equals milestone 3n's est to a median 1e-4 of f.
- H195 (the velocity explains the images). Robust cost of est_mu below est's in ≥ 75% of counted windows, separately for the Kinect (fr1/desk + fr1/xyz) and the Xtion (fr3).
- H196 (its timing carries it). est_mu's cost below mu_shuf's in ≥ 75% of windows, Kinect and Xtion separately.
- H197 (the velocity is part of what f compensates). Kinect pooled: median |e<sub>est_mu</sub>| ≤ 0.9 × median |e<sub>est</sub>| (milestone 3n: 2.55%).
- H198 (one rig, one offset). est_mud: robust sd (1.4826 MAD) of δ̂ across windows ≤ 5 ms in each of the three sequences, and |median δ̂| ≤ 10 ms (TUM's streams are time-stamped on one machine, Sturm et al. 2012).
- H199 (milestone 3o's bias is on real data). Kinect pooled: median |δ̂<sub>est_d</sub> − δ̂<sub>est_mud</sub>| ≥ 2 ms.
- H200 (with the velocity in, the offset is a nuisance for f). Kinect pooled: median |f<sub>est_mud</sub> − f<sub>est_mu</sub>|/f ≤ 0.3%.
- Reported, unscored. Per sequence and model: convergence, iterations, median and range of e and |e|, t̂<sub>r</sub>, δ̂ and the solver's sd; median scale s, mocap spread and speed; cost and rms against est; Spearman of each window's change in |e| (est_mu against est) with its median speed;
tumvel.png. - Caveat set in advance. The velocity is a central difference of mocap positions over 40 ms, and its scale comes from the window's own start, so a window with a short baseline gets a noisy s; the velocity is held at the camera stamp when δ̂ is free; mocap–optical-centre lever arm and rotation taken as TUM gives them; windows of one sequence share a scene and session; the board reference is a 13-view calibration (milestone 4n).
2026-11-18 · Milestone 3o result: a 1.2 s window estimates the gyro–camera time offset and its focal shift vanishes (20 ms late: 0.81% → 0.05%), but the offset lands a median 3 ms off with a claimed sd of 0.46 ms, because it soaks up the unmodelled linear velocity; four of six predictions held
- Run.
experiments/m3o_offset/run.py (30 windows × 21 solves, under 10 min on 4 cores) then analyse.py (summary.json, offset.png), as planned below. All 630 solves converged (median 14–16 iterations). tests/test_m3o.py (3 tests) passed before the run. Windows' median rms rate 0.30 rad/s, rms angular acceleration 3.5 rad/s<sup>2</sup>. - H189 refuted. known_d: median |δ̂ − δ| 2.8–3.6 ms at the five offsets (bar 2 ms); range −8.9 to +7.1 ms; median signed error −0.9 to −1.9 ms.
- H190 held. est_d: median |δ̂ − δ| 2.6–3.6 ms (bar 4 ms), median |t̂<sub>r</sub> − 33.3 ms| 3.0–3.3% (bar 10%); median t̂<sub>r</sub> 33.0 ms at every offset. The solver's t<sub>r</sub>–δ correlation is a median −0.01 to −0.02: readout and offset do not trade off.
- H191 held. δ = 20 ms: median |Δ| 0.050% with the offset free against 0.81% with it fixed at 0 (ratio 0.06; milestone 3m had 0.83% for its known model). At every offset the free models stay at 0.03–0.05%, against 0.14–0.81% fixed.
- H192 held. δ = 0: median |e| 1.76% (known_d) against 1.78% (known), 0.96% (est_d) against 1.01% (est). Freeing the offset costs nothing measurable when there is none.
- H193 refuted. The robust sd of (δ̂ − δ)/sd<sub>δ</sub> is 9.9 (150 known_d solves; bar [0.67, 1.5]): the solver claims 0.46 ms and is off by about 3.
- H194 held, barely. Spearman of |δ̂ − δ| (known_d, 10 ms) against rms angular acceleration −0.32 (bar ≤ −0.3).
- Unscored. est with the offset fixed at 0 drifts its readout to 31.7 ms at 20 ms (milestone 3m: 31.8 ms). Each window's offset error is nearly the same at every true offset (the spread in
offset.png moves rigidly with δ), so it is a per-window bias, not noise. - Post hoc diagnosis (
diag.py, diag.json; written after the scoring; known_d at 10 ms, first 12 windows, not a hypothesis test). Median |δ̂ − δ|: both noises on 2.96 ms (sd claimed 0.42); pixel noise only 3.09; gyro noise only 2.81; no noise at all 2.97 ms. With the generator's linear velocity set to zero: 0.00 ms without noise, 0.25 ms with both noises (claimed sd 0.33). With the linear velocity on and solved as a free velocity per keyframe (all 12 converged): 0.49 ms (rms 1.0, claimed sd 0.66). So the 3 ms is not noise: known_d and est_d assume u = 0, and the offset is the parameter that bends to mimic the rolling-shutter smear of the camera's translation (milestone 3k's part B velocities). The solver's sd is honest once the model is right. - Reading. A free gyro–camera offset is cheap and safe: it removes the focal shift a late gyro causes (a 20 ms offset's 0.81% falls to 0.05%), leaves the readout unbiased, and costs nothing at zero offset. But the offset estimate itself is only good to about 3 ms when the camera's linear velocity is left out of the model, and the solver reports a tenth of that. With a free velocity it is good to about half a millisecond. For TUM (the next step): the offset can go in free, but its value is only a calibration of the mocap timing if the linear velocity is modelled too, which on TUM did not converge in 300 iterations (milestone 3n); here, on synthetic windows, it converged in every case. For a practitioner: put the offset in the window as a nuisance parameter for f; do not report it as the rig's time offset unless the window also models translation.
- Caveat. Synthetic; milestone 3k's scenes and part B linear velocities, milestone 3m's smooth-rate model (0.5–3 Hz), one rate level; ω and u constant over each readout; no SIFT front end or outliers; gyro in the camera frame; the diagnosis uses 12 windows and one offset, and the free-velocity case was not run with the readout free.
2026-11-18 · Milestone 3o planned: can one 1.2 s window estimate the gyro–camera time offset, with the readout known and with it free, and does that remove a late gyro's focal shift?
Written before experiments/m3o_offset/run.py was first run in full and before analyse.py was written. Milestone 3n's next step is to estimate the mocap–camera offset on TUM, but the TUM hosts refuse the proxy again this session (curl CONNECT 403 for cvg.cit.tum.de, webshare.cvg.cit.tum.de, vision.in.tum.de), so this milestone builds and tests the estimator on milestone 3m's synthetic part T first, which is what the TUM run would then use. Seen before this plan: tests/test_m3o.py (3 tests) passed; one timing job (window 0, 21 solves, 34 s) whose results were deleted unread.
- Method (
experiments/m3o_offset/tsba.py, run.py). tsba.TWindow is milestone 3l's rsba.Window with the gyro as a 200 Hz signal g(s) and keyframe i's rate read by linear interpolation at t<sub>i</sub> − δ̂; δ̂ is one more camera-block parameter (Jacobian by central difference through the interpolation; with δ̂ fixed the solve equals rsba.Window's, tested). Windows: milestone 3m's part T (30 seeds, σ = 0.3 rad/s per axis as a sum of three sinusoids at U(0.5, 3) Hz, 1/30 s readout held over each keyframe, 0.5 px noise). Gyro: g(s) = w(s + δ) + white noise 0.1°/s per axis per sample (milestone 3l's level, one draw per seed shared by every δ), δ ∈ {−10, 0, 5, 10, 20} ms. Models: known (t<sub>r</sub> = 1/30 s, δ̂ = 0 fixed), est (t<sub>r</sub> free from 0, δ̂ = 0 fixed), known_d (t<sub>r</sub> known, δ̂ free from 0), est_d (both free from 0); plus clean (global shutter, no term). Scored: the paired shift Δ = (f − f<sub>ref</sub>)/f<sub>true</sub>, f<sub>ref</sub> the same model at δ = 0; e = (f − f<sub>clean</sub>)/f<sub>true</sub>; and δ̂ − δ. First-order size: at the windows' rms angular acceleration (about 3 rad/s<sup>2</sup>) a 1 ms offset changes the rate by about 1% of its rms, so a 1.2 s window of several thousand observations should pin δ to about a millisecond if the poses do not absorb it. - Counted. Every window; a solve that does not converge (status 0) is reported and left out of that model's figures.
- H189 (the offset is identifiable when the readout is known). known_d: median |δ̂ − δ| ≤ 2 ms at each of the five δ.
- H190 (offset and readout are jointly identifiable). est_d: median |δ̂ − δ| ≤ 4 ms at each δ, and median |t̂<sub>r</sub> − 33.3 ms| ≤ 10% of 33.3 ms at each δ.
- H191 (freeing the offset removes a late gyro's focal shift). δ = 20 ms: median |Δ<sub>known_d</sub>| ≤ 0.3 × median |Δ<sub>known</sub>| (milestone 3m: 0.83% for known).
- H192 (freeing it costs little when there is no offset). δ = 0: median |e<sub>known_d</sub>| ≤ median |e<sub>known</sub>| + 0.1 pp, and median |e<sub>est_d</sub>| ≤ median |e<sub>est</sub>| + 0.1 pp.
- H193 (the solver's sd of δ̂ is honest). known_d, pooled over δ: the robust sd (1.4826 MAD) of (δ̂ − δ)/sd<sub>δ</sub> ∈ [0.67, 1.5].
- H194 (identifiability follows the angular acceleration). known_d at δ = 10 ms: Spearman of |δ̂ − δ| against the window's rms angular acceleration at the keyframes ≤ −0.3.
- Reported, unscored. Per δ and model: convergence and iterations; median and range of Δ, e, δ̂ − δ, t̂<sub>r</sub>; the median t<sub>r</sub>–δ correlation from the solver's covariance (est_d); est's readout drift with δ fixed at 0 (milestone 3m: 31.8 ms at 20 ms);
offset.png. - Caveat set in advance. Synthetic; milestone 3k's scenes, milestone 3m's smooth-rate model with its chosen band, one rate level; ω constant over each readout; no SIFT front end or outliers; the gyro is in the camera frame (no rotation fault); linear interpolation of a 200 Hz signal; the offset is constant over the window.
2026-11-18 · Milestone 3n result: on real TUM video, motion-capture rates in the window solve find a rolling shutter on both cameras (Kinect readout 27.6 ms, Xtion 25.9 ms, positive in all 54 windows, about 0 with shuffled rates) and cut the median focal error from 2.87% to 1.88% (fr1/desk), 5.00% to 3.96% (fr1/xyz) and 7.56% to 3.63% (fr3); four of five predictions held
- Run.
experiments/m3n_tumrs/run.py on 64 of milestone 3b's windows (the 65th had no solution there), about 2 h on 4 cores (a first attempt was stopped by the 30 min background limit before writing anything; run.py now checkpoints each window and the rerun reproduced every readout it had printed), then analyse.py (summary.json, tumrs.png). 54 windows counted: fr1/desk 14, fr1/xyz 17, fr3/stf 23. Not counted: 3 with a mocap gap at a keyframe (fr1/desk 12, fr1/xyz 5, fr3 0), 6 whose nfba start did not converge, 1 with an undefined sd. Every counted window's none and est solve converged; shuf is missing in 3 windows whose source window had a mocap gap; est_u converged in only 6 of 54 (300-iteration cap), so its figures are not read. tests/test_m3n.py (4 tests, including the sign of the rate) passed before the run. - Check held. none's f equals milestone 3b's to a median 6e-6 of f; the nfba start equals it exactly in every window. In 8 of 54 windows this solver's tighter tolerance moves f by more than 1e-4, and in fr3 window 12 (an f already 37% off) by 10.7%.
- H184 held. fr1/desk, est: median t̂<sub>r</sub> 27.3 ms, positive in 14 of 14 windows (all 54 counted windows are positive, range 13.4–41.4 ms). Positive means rows exposed top to bottom, the usual CMOS readout.
- H185 held. fr1/desk robust sd of t̂<sub>r</sub> 3.0 ms, 11% of the median (bar 40%). The solver's own median sd is 0.67 ms, so window-to-window scatter is again about 4.5× what the solver claims.
- H186 held. With another window's rates, fr1/desk's median |t̂<sub>r</sub>| is 2.8 ms, 0.11× est's (bar 0.3), and est's cost is below shuf's in 13 of 13 fr1/desk windows that have a shuf solve (51 of 51 over all sequences). The fit uses the timing of the rates, not just one extra parameter.
- H187 held. fr1/desk median |e| 1.88% with the term against 2.87% without (0.66×, bar 0.7); |e| falls in 12 of 14 windows.
- H188 refuted. The slow Xtion sequence (median rms rate 0.18 rad/s) gains most: median |e| 7.56% → 3.63% (0.48×; predicted ≥ 0.8×), lower in 14 of 23 windows. Milestone 3k's bound used the window's mean rate over 1.1 s (0.08 rad/s); the per-keyframe rates the mocap gives are about twice that, and the window solve amplifies the shutter's distortion well past the first-order stretch (milestone 3k, H168), so that bound was too low.
- Unscored. fr1/xyz: 5.00% → 3.96% (lower in 12 of 17), readout 27.8 ms, robust sd 4.3 ms, shuf 0.9 ms. Pooled Kinect (31 windows): 27.6 ms, robust sd 4.0 ms; median |e| 4.12% → 2.55%. Xtion: 25.9 ms, robust sd 5.2 ms, shuf −0.5 ms. Sequence medians of f with the term: fr1/desk 523.7 px (+1.05% from the board), fr1/xyz 517.9 (−0.07%; without the term 511.3, −1.35%), fr3 544.2 (+2.54%; without 541.2, +1.97%). Cost falls a median 12% (fr1/desk), 10% (fr1/xyz), 14% (fr3) against none. The gain does not follow the window's rms rate (Spearman +0.11, −0.20, −0.04). Sensitivity to the rate's differencing span (d = 10 / 20 / 40 ms): Kinect readout 24.6 / 27.6 / 29.8 ms and |e| 2.92 / 2.55 / 2.36%; Xtion 19.3 / 25.9 / 36.4 ms and |e| 4.56 / 3.63 / 5.01%. The longer span smooths the rate and the readout grows to keep their product, so the Xtion's readout is known only to about ±30% from this data, and its f gain is largest at the 20 ms span used for scoring; the Kinect is stable to about ±10%.
- Reading. This is the programme's first real-data evidence for a cause of the natural-feature focal error that milestones 3b–3j could not explain. Both TUM cameras are rolling-shutter, with a readout of about 26–28 ms of their 33 ms frame period, and modelling it with the camera's own measured rotation removes a third (Kinect, hand-held desk) to a half (Xtion) of the median focal error of a 1.2 s window. The null control is clean: rates from the wrong time give a readout near zero and never fit better. What is left (1.9–4.0% median) is still well above what the solver's sd claims, so the shutter is part of the story, not all of it. For a practitioner: a robot with an IMU should put a rolling-shutter term in any short self-calibration window and estimate the readout there, as milestones 3l and 3m found on synthetic data; a target-free 1.2 s window measures it to about 3–5 ms per window on these sensors.
- Caveat. The readouts are estimates from motion capture, not a datasheet (no sourced figure for either sensor was found; Kerl et al. 2015 model both cameras' rolling shutter on TUM-style data but the text read here gives no readout); the Xtion figure depends on the rate's differencing span; mocap–camera timing and the optical-centre rotation are taken as TUM gives them; the free-velocity model did not converge in 300 iterations and was not tried longer; windows of one sequence share a scene; H188 was set from a bound that used mean rather than per-keyframe rates.
2026-11-18 · Milestone 3n planned: the TUM rolling-shutter test. With motion-capture angular rates in the window solve and the readout free, does TUM's Kinect show a rolling shutter, and is it what a 1.2 s window's focal length compensates?
Written before experiments/m3n_tumrs/run.py was first run on any window and before analyse.py was written. Milestones 3k–3m built and stress-tested the fix on synthetic windows and specified this test (per-keyframe mocap rates, readout free); it waited on the TUM host, which answers the proxy this session (curl 200 for cvg.cit.tum.de; the three sequences fetched by ../m3b_natural/fetch.sh). Neither the Kinect's nor the Xtion's colour readout time is sourced here, which is why it is estimated, not given. Seen before this plan: nothing from this milestone's solves; milestone 3m's synthetic result says a free readout absorbs a rate-scale error exactly and tolerates a 10 ms mocap–camera offset (0.35%) and a 10° rotation (0.42%).
- Method (
experiments/m3n_tumrs/run.py). Milestone 3b's windows (12 keyframes, every 3rd frame) of fr1/desk, fr1/xyz (Kinect) and fr3/structure_texture_far (Xtion). Per window: milestone 3b's SIFT tracks and nfba solve from the start that won there; from that solution, milestone 3l's rsba.Window (all five intrinsics free, soft-L1 LM) under none (t<sub>r</sub> = 0: milestone 3b's model in this solver), est (per-keyframe rates, t<sub>r</sub> free from 0), est_u (plus a free linear velocity per keyframe) and shuf (est with the rates of the window five on in the same sequence, circular: a null with the rates' size but not their timing). Rates: TUM's ground truth is the colour camera's optical-centre pose at 100 Hz (Sturm et al. 2012); the camera-frame body rate at a keyframe's timestamp is log(R(t−d)<sup>T</sup>R(t+d))/2d with R slerped, d = 20 ms, entered as w = −ω<sub>b</sub> (tests/test_m3n.py checks the sign against a camera turning at a known rate), so t̂<sub>r</sub> > 0 means rows exposed top to bottom. Sensitivity, unscored: d = 10 and 40 ms. e = (f − f<sub>ref</sub>)/f<sub>ref</sub> against milestone 4n's board calibration (fr1 518.28, fr3 530.71 px). - Counted. Milestone 3b's counting rule (converged, ≥ 100 tracks, inlier rms < 1.5 px, 300 < f < 1000, defined sd) on the none solve, plus mocap rates at every keyframe; a model's solve that does not converge is reported and left out of that model's figures.
- Check, not a hypothesis. none's f equals milestone 3b's (median |Δf|/f ≤ 1e-4); reported either way.
- H184 (the Kinect has a rolling shutter a window can see). fr1/desk (median mocap rate 0.38 rad/s, the fastest), est: median t̂<sub>r</sub> ∈ [10, 45] ms and t̂<sub>r</sub> > 0 in ≥ 75% of counted windows.
- H185 (one sensor, one readout). fr1/desk, est: robust sd (1.4826 MAD) of t̂<sub>r</sub> across windows ≤ 40% of its median.
- H186 (the timing carries it, not the extra parameter). fr1/desk: median |t̂<sub>r</sub>| under shuf ≤ 0.3 × that under est, and est's robust cost is below shuf's in ≥ 75% of windows.
- H187 (the shutter is part of what the Kinect's f compensates). fr1/desk: median |e<sub>est</sub>| ≤ 0.7 × median |e<sub>none</sub>|.
- H188 (it is not what the slow Xtion's f compensates). fr3/stf (median rate 0.08 rad/s): median |e<sub>est</sub>| ≥ 0.8 × median |e<sub>none</sub>| (milestone 3k's bound: its rates imply about a tenth of its error).
- Reported, unscored. Per sequence and model: counted windows, convergence, median and range of e and |e|, median t̂<sub>r</sub>, its sd and robust spread, cost and rms change against none; fr1/xyz under every hypothesis's statistic; est_u; d = 10 and 40 ms; the pooled per-unit t̂<sub>r</sub>; Spearman of each window's change in |e| against its rms rate;
tumrs.png. - Caveat set in advance. The mocap–camera time offset and the optical-centre rotation in TUM's ground truth are taken as given (milestone 3m: 10 ms and 10° cost under 0.5%); the rate is a central difference over 40 ms, so vibration faster than that is not in it; rates are constant over a readout; the windows of one sequence share a scene and session; the board reference is itself a 13-view calibration (milestone 4n), 0.3% from TUM's published one.
2026-11-18 · Milestone 3m result: the gyro fix survives a 10° camera–IMU rotation (0.42% median focal shift) and a 10 ms time offset (0.35%), but a gyro scale or readout 5% off with the readout held fixed costs 1.23%, always the same sign; a free readout absorbs it exactly; four of six predictions held
- Run.
experiments/m3m_gyrofault/run.py (30 windows × 26 solves, about 6 min on 4 cores) then analyse.py (summary.json, gyrofault.png), as planned below. All 780 solves converged except part T's plain solve in 1 of 30. tests/test_m3m.py (5 tests: the fault generators, the scale–readout identity, the scoring) passed before the scored analysis. Check passed: the exact-gyro clean, none, known and est solves equal milestone 3l's to 0 (the run is deterministic). - H178 held. 1° rotation, readout known: median |Δ| 0.044% (range −0.25% to +0.18%).
- H179 held. 10°: median |Δ| 0.42% (range −3.2% to +1.2%); 3° 0.15%; ratio 10°/1° 9.4.
- H180 held. With the readout free a scale error is absorbed to solver precision: |Δ| ≤ 4e-10 at s = 1.02 and 1.05, s·t̂<sub>r</sub>(s)/t̂<sub>r</sub>(1) = 1 to 1e-10 (t̂<sub>r</sub> 32.8 and 31.8 ms).
- H181 refuted. s = 1.05, readout known: median |Δ| 1.23% against the [0.05, 0.4]% predicted, and negative in all 30 windows (−0.26% to −3.4%); s = 1.02: 0.50%, also negative in all 30. The model sees ω only through ωt<sub>r</sub> (tested), so this is also the cost of a readout taken 2% or 5% too long. The first-order guess (relative error × milestone 3k's 4.17%) is 5× too small, and unlike milestone 3k's shift its sign does not depend on the window; why was not examined, and a readout too short (s < 1) was not run.
- H182 held. Time offset, readout known: median |Δ| 0.35% at 10 ms, 0.83% at 20 ms (2 ms 0.05%, 5 ms 0.14%). The offset's rate error is 2.4%, 6.1%, 12% and 24% of the rms rate; Spearman of |Δ| against each window's relative rate error only +0.21 to +0.35, so the window's geometry again sets most of the size.
- H183 refuted. A free readout does reduce the misalignment's shift (10°: 0.29% against 0.42%; 3°: 0.07% against 0.15%) and slightly the 20 ms offset's (0.71% against 0.83%; at 2–10 ms est is 0.05–0.2 pp worse). Its readout stays at 33.0 ms (10°) and drifts to 31.8 ms at 20 ms.
- Unscored. Part T's own baselines (smooth sum-of-sines rates, exact gyro): none 3.31% (29 of 30 converged), known 1.75%, est 1.00%: on these smoother paths the free readout beats the known one even with a perfect gyro, which part B did not show (1.13% against 1.14%); not examined. Against the clean f, the worst fault tested (20 ms, known) leaves |e| 1.91%, still under part T's none 3.31%.
- Reading. Of the three ways a real camera–IMU rig is wrong, the one that matters is the product of the gyro's scale and the readout time. Held at a spec-sheet value 5% off, it biases f by about a quarter of the error, always the same way, which pooling windows (milestone 3d) would not average out. Left free, the window absorbs it exactly. Misalignment and timing are forgiving: a 3° rotation or a 5 ms offset costs under 0.15%, below the 1.1% the window's own noise leaves, and even 10° or 20 ms stays under 1%. For a practitioner: estimate the readout in the window rather than importing it, align the gyro to a few degrees and sync it to within 10 ms. For the TUM test: motion-capture rates with a few degrees of body-to-camera rotation error and under 10 ms of timing error are good enough, provided the readout is free (TUM's actual mocap–camera timing error is not sourced here).
- Caveat. Synthetic, milestone 3k's scene and path models, no SIFT front end or outliers; ω constant over each readout; part T's sinusoid band (0.5–3 Hz) is chosen, not measured; one rate level; fault sizes swept, none from a datasheet; the faults were not estimated, only tolerated; the one-sign scale result is unexplained and only s > 1 was run.
2026-11-18 · Milestone 3m planned: does milestone 3l's gyro fix survive a wrong gyro (a camera–IMU rotation off by 1–10°, a 2–5% scale error, a 2–20 ms time offset)?
Written before experiments/m3m_gyrofault/run.py was first run in full and before analyse.py was written. Milestone 3l's caveat: its estimator was given the generator's own gyro model, so it showed the rolling shutter is identifiable, not that the fix survives the faults a real camera–IMU rig has. The TUM test (per-keyframe mocap rates, readout free) is still blocked: curl CONNECT 403 today for cvg.cit.tum.de, webshare.cvg.cit.tum.de and vision.in.tum.de. Mocap rates on TUM would themselves carry a time offset and a body-to-camera rotation error (the mocap markers are on the Kinect's body, not its optical centre), so this milestone also says how good those rates must be. Seen before this plan: one timing job (the first part B window, all 26 solves converged in 47 s); no focal length was printed.
- Method (
experiments/m3m_gyrofault/run.py). Milestone 3l's estimator (rsba.Window, all five intrinsics free, its start), milestone 3k's part B windows at σ = 0.3 rad/s (30 seeds, 1/30 s readout, 0.5 px noise) and milestone 3l's gyro noise draw. Each window is solved with the readout known (known) and free from 0 (est) under a gyro that is R**otated, ω<sub>meas</sub> = R(θ)ω, θ ∈ {1, 3, 10}° about a random axis per window; **S**caled, ω<sub>meas</sub> = sω, s ∈ {1.02, 1.05}; or **T**ime-offset, ω<sub>meas</sub>(t<sub>k</sub>) = ω(t<sub>k</sub> + δ), δ ∈ {2, 5, 10, 20} ms. Part T needs ω as a function of time, so its windows keep milestone 3k's scenes and linear velocities with ω(t) per axis a sum of three sinusoids at U(0.5, 3) Hz, random phases, sd 0.3 rad/s per axis (part B's σ), held constant over each keyframe's readout as milestone 3k's generator does. None of these fault sizes is from a datasheet or a calibration report; they are swept. Scored quantity: the paired shift **Δ = (f<sub>fault</sub> − f<sub>exact</sub>)/f<sub>true</sub> against the same window, pixel noise and model with the exact gyro, which removes the 0.3–1.4% estimation noise milestone 3l's post hoc found. First-order expectation: Δ ≈ (relative rate error) × milestone 3l's none shift (4.17%). - Check, not a hypothesis. The exact-gyro clean, none, known and est solves of parts R and S equal milestone 3l's (median |Δf|/f ≤ 1e-6).
- H178 (calibration-grade alignment is harmless). θ = 1°, known: median |Δ| ≤ 0.1%.
- H179 (misalignment grows about linearly and stays below the noise). θ = 10°, known: median |Δ| ≤ 1.0%, and median |Δ|(10°) / median |Δ|(1°) ∈ [5, 20].
- H180 (a free readout absorbs a scale error exactly). S, est: median |Δ| ≤ 0.01% at both scales, and median s·t̂<sub>r</sub>(s)/t̂<sub>r</sub>(1) within 0.5% of 1. (The model sees ω only as the product ωt<sub>r</sub>; a failure would be a solver artefact.)
- H181 (with the readout known, a scale error costs about its own size times the shutter's shift). S = 1.05, known: median |Δ| ∈ [0.05, 0.4]%.
- H182 (a time offset of 10 ms is tolerable). T, known: median |Δ| ≤ 0.6% at δ = 10 ms and ≤ 1.2% at 20 ms.
- H183 (a free readout does not rescue a rotated or late gyro). median |Δ<sub>est</sub>| ≥ 0.8 × median |Δ<sub>known</sub>| at θ = 10° and at δ = 20 ms (both must hold).
- Reported, unscored. Per fault and model: median and range of Δ and of e against the clean f; median t̂<sub>r</sub>; the rms rate error per δ and |Δ| against relative rate error (Spearman); part T's none and known |e| (its own baseline); convergence;
gyrofault.png. - Caveat set in advance. Synthetic, milestone 3k's scene and path models, no SIFT front end or outliers; ω constant over each readout (no vibration within a frame); the sinusoid bandwidth of part T is chosen, not measured; misalignment and offset are fixed per window and unmodelled by the estimator (estimating them is the next step if they matter); one rate level (σ = 0.3 rad/s).
2026-11-17 · Milestone 3l result: with a gyro, a rolling-shutter term in the window solve cuts milestone 3k's hand-held focal shift from 4.17% to 1.14% (0.71% with a free velocity), and one window estimates the readout to 3.6% with no false shutter on global-shutter data; five of six predictions held
- Run.
experiments/m3l_gyro/run.py (76 windows, 9 models on part B σ = 0.3, 5 elsewhere; under 20 min on 4 cores) then analyse.py (summary.json, gyro.png), as planned below. Every solve converged except milestone 3k's plain model at σ = 1.0 (22 of 30). Check passed: this solver's clean and none f equal milestone 3k's nfba solves to a median 1.6e-6 and 4.2e-5 of f (8 of 76 none solves differ by more than 1%, the diverging σ = 1.0 ones). - H172 refuted, narrowly. Part B σ = 0.3: median |e| 1.14% with the gyro and known readout against 4.17% without (ratio 0.27; bar 0.25). Range −6.2% to +5.2% (none: −18% to +14%).
- H173 held. Part A, pitch and yaw at 1 rad/s: 0.29% (none: 5.9%); pitch 0.47%, yaw 0.21%.
- H174 held. A free linear velocity per keyframe: 0.71% against 1.14%; median signed e 0.03% (known: 0.14%).
- H175 held. Readout free from 0: median t̂<sub>r</sub> 33.4 ms (range 30.6–37.8; true 33.3), median relative error 3.6%, median sd 0.23 ms; median |e| 1.13% (bar 2.28%). With u free too: 33.3 ms, 1.2%, |e| 0.80%.
- H176 held. Global-shutter rendering solved with a free readout: |t̂<sub>r</sub>| ≤ 3.3 ms in 30 of 30 windows (range −0.33 to +0.39 ms); f moves a median 0.01%.
- H177 held. 1 °/s gyro bias: 1.15% against 1.14% (0.01 pp); 5 °/s: 1.10%.
- Unscored. σ = 1.0: none 18.9% (9 of 30 over 50%), known 1.0%, est 0.69% (readout 33.2 ms, 1.0% error), all converged. Part A with a free readout: yaw 32.6 ms (2.1% error) but pitch 24.2 ms (27% error, range 16–33) and f 0.99%; the null solve at constant pitch finds −6.7 ms on global-shutter data. A steady pitch rate makes the rolling shutter a vertical stretch that f, c<sub>y</sub> and k can partly absorb, so the readout is weakly identified; random hand-held motion separates them. The remaining shift is 2–3 solver sd (known_u 2.2, est 3.0), so the solver's own sd still understates it.
- Post hoc (
explore.py, explore.json; written after the verdicts above were read). Six part B σ = 0.3 windows with the full model (gyro, known readout, free u): with the pixel noise off, f is recovered to under 1.4e-7 of f in all six, so the model and solver are exact and the 0.3–1.4% left with 0.5 px noise is estimation noise on a 36-parameter-larger model, not a bias. The two renderings share 99.2–99.7% of their observations. - Reading. Given a gyro, a rolling shutter is a solvable nuisance in a 1.2 s target-free window, not a hard floor: the shift falls by about 4× with the readout known and stays down when the window estimates the readout itself, which it does to a few percent from ordinary hand-held motion while finding no shutter on a global-shutter camera. A 5 °/s gyro bias did not matter here: its rotation over one readout (at most 2.9 mrad per axis) is small beside that of the 0.3 rad/s rates (about 10 mrad), and why it moves f even less than that ratio suggests was not examined. For a practitioner: put the gyro in the window solve, leave the readout free, and avoid steady single-axis calibration motions. For the programme: the TUM test is now specified (per-keyframe mocap rates, readout free); a readout near zero rules the shutter out as what TUM's f compensates, a consistent one rules it in and measures the Kinect's readout.
- Caveat. Synthetic, milestone 3k's one scene and path model, no SIFT front end or outliers; the estimator's shutter model equals the generator's (constant ω and u over one readout), so this shows identifiability and the size of the fix, not robustness to a wrong model (vibration within a frame, a gyro time offset, a wrong camera–IMU rotation); the gyro noise and bias are not from a datasheet.
2026-11-17 · Milestone 3l planned: with a gyro, does a rolling-shutter term in the 1.2 s window solve remove milestone 3k's focal shift, and can one window estimate the readout time itself?
Written before experiments/m3l_gyro/run.py was first run in full and before analyse.py was written. Milestone 3k found that a 1/30 s rolling shutter moves a short window's f by a median 4.2% under 0.3 rad/s hand-held rates, with a sign set by the window's geometry, so no mean-rate correction can undo it. The fix a robot would use is to model the shutter in the solve with the angular velocity its IMU already measures (Karpenko et al. 2011; Hedborg et al. 2012; Oth et al. 2013 estimate the readout with a target). This milestone asks whether that fix works inside the project's own window, and whether the readout time is identifiable from one target-free 1.2 s window, which is what the TUM test (per-frame mocap rates, readout unknown) would need. The TUM host still refuses the proxy (connection failed for cvg.cit.tum.de, webshare.cvg.cit.tum.de, vision.in.tum.de). Seen before this plan: one timing job (part B, σ = 0.3, the first seed): every model below converged in 11–42 iterations, about 20 s; its focal lengths were not printed.
- Method (
experiments/m3l_gyro/). rsba.py: milestone 3b's window model plus a rolling-shutter term, a point seen at row v exposed at τ = t<sub>r</sub>(v − H/2)/H, x<sub>cam</sub> = exp([ω<sub>i</sub>τ]<sub>×</sub>)(R<sub>i</sub>X + t<sub>i</sub>) + u<sub>i</sub>τ, ω<sub>i</sub> the keyframe's gyro reading, solved by the same robust Schur-complement Levenberg–Marquardt as nfba.bundle (soft-L1 at 1 px, all five intrinsics free, milestone 3k's start). Data: milestone 3k's own windows, motion and noise draws (part B, both σ; part A pitch and yaw at 1 rad/s), rendered at t<sub>r</sub> = 1/30 s. Models: clean (global-shutter rendering, no term; the reference f), none (rolling-shutter rendering, no term: milestone 3k's solve in this solver), known (gyro ω, t<sub>r</sub> = 1/30 s given, u = 0), known_u (plus a free u per keyframe), est (t<sub>r</sub> free from 0, u = 0), est_u (t<sub>r</sub> and u free), bias1 and bias5 (known, with a gyro bias of 1 and 5 °/s per axis, random sign, fixed over the window), null (global-shutter rendering solved with est). Gyro white noise 0.1 °/s per axis per keyframe. The bias and noise are not from a datasheet (the Bosch and TDK product pages were unreachable from this container), so the bias is swept. e = (f − f<sub>clean</sub>)/f<sub>true</sub>; known_u, est_u, bias1, bias5 run on part B σ = 0.3 only. tests/test_m3l.py (4 tests: zero τ equals nfba.project; at the generator's own row times the model equals milestone 3k's camera to 1e-6 px; known shutter recovers f within 0.3% on a small noiseless-ish window; a free t<sub>r</sub> is recovered within 10% there) passed before this plan. - Check, not a hypothesis. This solver's clean and none f should match milestone 3k's nfba solves of the same windows (median |Δf|/f ≤ 0.1%); reported either way.
- H172 (a gyro and a known readout remove the shift). Part B σ = 0.3: median |e<sub>known</sub>| ≤ 0.25 × median |e<sub>none</sub>|.
- H173 (rotation only: near exact). Part A pitch and yaw at 1 rad/s pooled: median |e<sub>known</sub>| ≤ 0.5%.
- H174 (translation is the residue). Part B σ = 0.3: median |e<sub>known_u</sub>| < median |e<sub>known</sub>|.
- H175 (the readout is identifiable from one window). Part B σ = 0.3, est: median |t̂<sub>r</sub> − t<sub>r</sub>|/t<sub>r</sub> ≤ 0.2 and median |e<sub>est</sub>| ≤ 2 × median |e<sub>known</sub>|.
- H176 (no shutter where there is none). null, part B σ = 0.3: |t̂<sub>r</sub>| ≤ 0.1 × (1/30 s) in ≥ 90% of windows.
- H177 (a 1 °/s gyro bias does not matter). Part B σ = 0.3: |median |e<sub>bias1</sub>| − median |e<sub>known</sub>|| ≤ 0.3 percentage points.
- Reported, unscored. Per model and part: median and range of e, convergence, sd<sub>f</sub>, the shift in sd<sub>f</sub> units; t̂<sub>r</sub> and its sd for est, est_u and null; bias5; σ = 1.0 (where milestone 3k's solves diverged in 9 of 30) for none, known and est;
gyro.png. - Caveat set in advance. Synthetic, milestone 3k's scene and path model, no SIFT front end or outliers; the estimator's shutter model is the generator's (constant ω and u over one readout), so a pass shows identifiability and the size of the fix, not robustness to a wrong shutter model; the gyro's extrinsic rotation and time offset are taken as known; constant ω per readout ignores vibration within a frame.
2026-11-17 · Milestone 3k result: a rolling shutter can put a TUM-sized error into a 1.2 s window's focal length (random 0.3 rad/s hand-held rates, 1/30 s readout: median shift 4.2%), but its sign and size depend on the window's geometry, so the window's mean rotation rate does not bound it (R<sup>2</sup> 0.03); two of five predictions held
- Run.
experiments/m3k_shutter/run.py (156 jobs, each a global- and a rolling-shutter solve of the same scene and noise; about 6 min on 4 cores) then analyse.py (summary.json, shutter.png), as planned below. Every part A solve and every σ = 0.3 part B solve converged; at σ = 1.0, 24 of 30 converged and 9 of 30 shifted f by more than 50% (−80% to +5000%: diverged). The global-shutter solves land a median 0.31% from the true f. tests/test_m3k.py (4 tests) passed before the run. - H167 held. Pitch: median shift −1.1%, −2.5%, −4.5%, −7.3% at 0.25, 0.5, 1, 2 rad/s; ratios 2:1 = 1.64, 0.5:1 = 0.56. Roughly linear up to 1 rad/s, flattening at 2.
- H168 held. Pitch at 1 rad/s: median |shift| 6.4% against 1.3% from the isotropic first-order stretch (4.8×). The window's solve amplifies a vertical stretch it cannot represent, and the sign is not the stretch's: per seed −8.7% to +10.9% (2 of 8 positive).
- H169 refuted. Yaw moves f as much as pitch (median |shift| at 1 rad/s 5.6%, 0.88× pitch's; median sign positive, per seed −10.7% to +13.9%); only roll is small (1.2%, 0.19×). Yaw's shear is not absorbed by c<sub>x</sub> alone in a short window.
- H170 refuted. Hand-held random rates: the shift does not follow the window's mean pitch rate (OLS slope −96, R<sup>2</sup> 0.03, both σ pooled, the diverged solves included as planned). Post hoc, converged solves only, a regression on all three mean rates explains R<sup>2</sup> 0.18 (σ = 0.3) and 0.13 (σ = 1.0).
- H171 refuted. At each TUM window's mocap mean rate (fr1/desk 0.38, fr1/xyz 0.14, fr3/stf 0.08 rad/s median) with all of it in pitch, the implied shift reaches half the observed error in 10 of 57 windows (18%; bar 50%); median implied 0.53% against observed 5.25%. Explaining each window's error that way would take 1.8× (fr1/desk), 8.4× (fr1/xyz) and 20× (fr3/stf) its mocap mean rate.
- Unscored. Part B, σ = 0.3 rad/s per axis with 0.3 m/s translation rates: median |shift| 4.2%, sd 7.2%, range −18% to +14%; σ = 1.0: median 19%. The shift is 4–39 global-shutter sd<sub>f</sub> (pitch and yaw, 0.25–2 rad/s), so a window's own covariance never flags it. k<sub>1</sub> moves with f under pitch (−0.08 at 1 rad/s).
- Post hoc (
explore.py, explore.json; written after the verdicts above were read, not pre-registered). Rotation alone, σ = 0.1 rad/s: median |shift| 0.98% (13% of windows over 2%); σ = 0.3: 2.7% (67% over 2%); translation alone at 0.3 m/s: 1.0% (20% over 2%). The rotation term dominates; translation adds about a percent. - Reading. The bound this milestone set out to compute does not exist in the form planned: a rolling shutter's effect on a short window's f is not a small, sign-stable stretch scaled by the mean rate. Pitch and yaw rates both move f, by about 1.3× the rotation during readout (in radians), with a sign set by the window's geometry, and with hand-held random rates the effect is as large as TUM's observed error (4.2% at 0.3 rad/s per axis and the longest readout; 1% at 0.1 rad/s). So the rolling shutter is not ruled out as what the natural-feature f compensates, and the earlier finding that rotation does not predict error (milestone 3e, Spearman +0.04, again here) is no evidence against it, because the mean rate is the wrong predictor. Against it: fr3/stf, the slowest sequence (0.08 rad/s mean), has the largest error (7.2%), which random rates near its mean would not produce (0.1 rad/s gives about 1%) unless its high-frequency jitter or readout differ. The test that settles it needs the per-frame angular velocity (TUM's motion capture runs at 100 Hz) and a rolling-shutter term in the window model, with the readout time from a source; both wait for the TUM host.
- Caveat. Synthetic: one scene shape and path model, one camera, no SIFT front end or outliers; the readout is the 1/30 s upper bound, not a sourced Kinect figure; part A's 8 seeds per cell give rough medians; the TUM side bounds with the window's largest rotation, not its per-frame rates.
2026-11-17 · Milestone 3k planned: can a rolling shutter put a 5% error into a 1.2 s window's focal length? A synthetic sweep through the project's own solver, set against the TUM windows' motion-capture rates
Written before experiments/m3k_shutter/run.py was first run in full and before analyse.py was run. Milestones 3f–3j ruled out the static intrinsics (distortion, principal point) as what a natural-feature window's f compensates; the README's next candidates are a rolling shutter, track bias and longer windows. The TUM host refuses the proxy again in this session (curl CONNECT 403 for cvg.cit.tum.de; webshare and vision.in.tum.de time out), so the track-bias re-track and longer windows wait, and this milestone asks how large the rolling-shutter effect can be from first principles and the motion already stored. Which camera TUM's Kinect units use and their readout time are not sourced here, so the readout is set to the whole 30 Hz frame period, 1/30 s, an upper bound for any 30 Hz rolling shutter; the effect is linear in rate × readout to first order, so results scale down for a shorter readout. Seen before this plan: one pilot job (part A, pitch at 1 rad/s, seed 0): the rolling-shutter solve moved f by +10.9% while the isotropic first-order stretch model below predicts −1.3%; the hypotheses are set knowing that one number.
- Method (
experiments/m3k_shutter/). rs.py: a rolling-shutter camera, row v exposed at τ = t<sub>r</sub>(v − c<sub>y</sub>)/H, camera moving with angular velocity ω and linear velocity u (camera frame) during readout, solved as a fixed point (10 iterations, under 1e-6 px). Synthetic windows shaped like milestone 3b's: 12 keyframes over 1.1 s on a smooth hand-held path (15° rotation, 0.25 m translation, small jitter), 400 points 1.5–5 m away, 0.5 px noise, fr1 board intrinsics (f 518.28, c (315.7, 260.1), k<sub>1</sub> 0.25, k<sub>2</sub> −0.75). Each window is solved twice with milestone 3b's nfba.bundle (all five intrinsics free, start f × 1.08, centred c, k = 0, the start milestone 3h's tests use): once rendered global-shutter, once rolling-shutter, same scene and noise; e = (f<sub>rs</sub> − f<sub>gs</sub>)/f<sub>true</sub> is the shutter's own shift. Part A: constant ω about one camera axis (pitch x, yaw y, roll z) at 0.25, 0.5, 1, 2 rad/s, 8 seeds each. Part B: hand-held, each keyframe's ω drawn N(0, σ<sup>2</sup>) per axis, σ ∈ {0.3, 1.0} rad/s, u N(0, 0.3<sup>2</sup>) m/s per axis, 30 seeds each. The isotropic first-order model (rs.first_order_bias): a pitch rate stretches rows by s = ωt<sub>r</sub>f/H, and an isotropic f fitted by least squares takes the vertical share of it, e<sub>iso</sub> = −s Σdy<sup>2</sup>/Σ(dx<sup>2</sup> + dy<sup>2</sup>). The gain g = median e/(ωt<sub>r</sub>) for pitch at 1 rad/s. tests/test_m3k.py (4 tests: zero readout equals nfba.project; the fixed point converges; each point's row time matches its row; a pitch rate is a vertical stretch of the predicted size) passed before this plan. - H167 (the shift is linear in rate). Pitch, median e at 2 rad/s over median e at 1 rad/s in [1.6, 2.4], and at 0.5 over 1 in [0.4, 0.6].
- H168 (the window amplifies the stretch). Pitch at 1 rad/s: median |e| ≥ 2× |median e<sub>iso</sub>|. (The pilot suggests far more, and of the opposite sign.)
- H169 (pitch is the axis that matters). At 1 rad/s, median |e| for yaw and for roll each ≤ 0.5× pitch's.
- H170 (hand-held: the mean pitch rate predicts the shift). Part B, both σ pooled: ordinary least squares of e on g · t<sub>r</sub> · (mean over keyframes of ω<sub>x</sub>) has slope in [0.5, 1.5] and R<sup>2</sup> ≥ 0.5.
- H171 (a rolling shutter is a plausible cause on TUM). For milestone 3c's 57 scored windows, take each window's mean rate ω̄ = milestone 3e's motion-capture rotation (largest between any two keyframes) / 1.1 s, put all of it in pitch at t<sub>r</sub> = 1/30 s: the implied |e| = |g| ω̄ t<sub>r</sub> is ≥ half the observed |e| in ≥ 50% of windows. This is generous to the rolling shutter (all rotation in pitch, constant sign, the longest readout); a refutation says it cannot be the main cause even then.
- Reported, unscored. Per axis and rate: median e, g, the shift in units of the global-shutter sd<sub>f</sub>, the k<sub>1</sub> shift; part B: sd and median |e| per σ; per TUM sequence: median observed |e|, mocap rate, implied |e|, and the rate that would be needed to explain |e| (and its ratio to the mocap rate); Spearman of rate against |e| (milestone 3e found +0.04 with the rotation itself);
shutter.png. - Caveat set in advance. Synthetic: one scene shape, one path model, one camera; the global-shutter solve itself is not the TUM solve (no SIFT front end, no outliers). The TUM side has no per-keyframe pitch rate (only the window's largest rotation is stored; the video and motion capture are refused today), so H171 can only bound, not attribute; the sign of TUM's error cannot be tested against the sign of the pitch here. Hand-held jitter faster than the 0.1 s keyframe spacing is not in the mocap rate and could make instantaneous rates larger than ω̄.
2026-11-17 · Milestone 3j result: re-solved on the TUM video, the board's distortion is worse than first order said (median focal error 5.25% to 10.5%, f above the board's in 52 of 53 windows), the principal point mildly worse (5.87%), all four 10.7%; first order gets the size right on median but the sign wrong in a quarter of the distortion re-solves
- Run.
experiments/m3j_tumpin/run.py (about 12 min on 4 cores after a 1.3 GB download) then analyse.py (summary.json, tumpin.png). All 57 windows rebuilt; 4 of 57 k-pinned re-solves hit the iteration cap and are excluded (fr1/desk 16, fr1/xyz 2, 13, 16); every c and ck re-solve converged with its pins held. tests/test_m3j.py (4 tests, the scoring on known cases) passed before the scored analysis. - Deviation from the plan. The plan said the free f would be asserted equal to milestone 3c's. 51 of 57 rebuilt windows match it to 1e-6, 55 to 0.1%; fr3/stf 0 differs by 0.18% and fr3/stf 12 by 3.35% (this session's OpenCV SIFT build). Instead of failing the run,
analyse.py scores every hypothesis on all 57 (the free solve of this run as the baseline for each pin) and again on the 55 reproduced windows; every verdict is the same on both (summary.json, reproduced_only). - H161 held. k pinned: median |e<sub>p</sub>| 10.46% against 5.25% free (first order from the fresh covariances 8.37%; milestone 3g's 8.72%; 3h's shrunk transfer 7.12%). The re-solve is worse than every earlier estimate. f ends above the board's f in 52 of 53 windows. Per sequence: fr1/desk 11.3%, fr1/xyz 18.1%, fr3/stf 6.1% (free 2.80%, 5.18%, 7.24%).
- H162 held, narrowly. c pinned: median |e<sub>p</sub>| 5.87% against 5.25% (first order 6.75%). The sign differs by sequence: fr1/desk 2.80% to 3.40%, fr1/xyz 5.18% to 7.10%, fr3/stf 7.24% to 5.88%.
- H163 held. All four pinned: 10.74% (first order 7.22%).
- H164 refuted. First order and re-solve moved f the same way in 74% (k), 95% (c) and 70% (ck) of windows; bar 90%. The k disagreements are not only tiny steps: their median first-order step is 3.5% of f (agreements 9.2%), and among steps of 2% or more 78% agree.
- H165 held. Median re-solved/first-order step ratio 0.91 (k), 1.01 (c), 0.87 (ck), inside [0.67, 1.5]; fr1/xyz's k ratio 0.67 sits on the edge and its ck ratio is 0.43.
- H166 held. k pinned: median relative gap between the re-solved and first-order sd ratios 0.078 (bar 0.2).
- Unscored. Median robust-cost rise when pinned: k 3.2% (fr1/desk 5.1%, fr1/xyz 6.6%, fr3/stf 0.4%), c 1.2%, ck 5.3% (fr1/xyz 11.8%). The step-binned ratio is noisy below a 2% step (single windows with ratios from −256 to +158 where the first-order step is near zero) and 0.9–1.0 for k at 2–10%. One window (fr3/stf 19) has a first-order step of ±6× f, a degenerate covariance, against a re-solved +27%; it is off the figure's scale.
- Reading. Re-solving confirms the first-order verdicts of milestones 3g and 3i and makes the distortion one sharper. The fr1 Kinect's natural-feature windows do not accept the board's distortion: pinning it raises the robust cost 5–7% and drives f up by 10–18%, above the board's own f in nearly every window. That is not a linearisation artefact; on TUM it is larger than first order predicted. fr3, whose images TUM ships already undistorted, barely resists the pin (0.4% cost) and its error barely moves. So the windows' free k and f are a matched pair: the board's k is the wrong partner for any f the window supports. The principal point is a weaker, sequence-dependent lever: it helps the far, textured fr3/stf scene and hurts both fr1 ones. For a robot, the practical rule from 3f–3i now holds on the re-solve: in a 1.2 s natural-feature window, leave the intrinsics free and pool windows (milestone 3d); do not import the board's distortion. The question of what the free f compensates stays open (rolling shutter, motion, track bias); the TUM video is now on disk, so those are testable.
- Caveat. 57 windows, three sequences, two Kinect units; the TUM reference is the board calibration the error is scored against; each pinned solve starts at the free solution and may stop in a nearby local minimum; 2 windows did not reproduce 3c's free f exactly (verdicts unchanged without them).
2026-11-17 · Milestone 3j planned: the TUM re-solve. Do the first-order verdicts of milestones 3g and 3i survive when each natural-feature window is re-solved with the board's distortion, principal point, or both pinned?
Written before experiments/m3j_tumpin/run.py was first run and before any pinned TUM solve was computed. The TUM host answers the proxy again today (cvg.cit.tum.de redirects to webshare.cvg.cit.tum.de, 200, 448 MB for fr1/xyz), so the re-solve that milestones 3g, 3h and 3i deferred can run. Seen before this plan: nothing from the TUM video beyond the downloads; the board references are 3i's.
- Method (
experiments/m3j_tumpin/run.py). Milestone 3c's 57 scored windows, each rebuilt exactly as 3c did (3b's pipeline from its lowest-cost start, 3c's OpenCV RNG seed, then 3c's refit); the analysis asserts the free f equals 3c's. From that solution, milestone 3h's bundle_pinned re-solves with k = (k<sub>1</sub>, k<sub>2</sub>), c = (c<sub>x</sub>, c<sub>y</sub>) and ck = all four pinned at the board's Zhang k<sub>1</sub> + k<sub>2</sub> values (milestone 3i's board_ref, asserted equal to 3i's). First order f<sub>c</sub> from the fresh covariance with milestone 3g's formula. Error e = (f − f<sub>ref</sub>)/f<sub>ref</sub> against milestone 3c's per-window reference, as in 3c–3i. - H161 (pinning the board's distortion hurts, re-solved). Median |e<sub>p,k</sub>| ≥ the free median |e| (3g first order 8.72% against 5.25%; 3h's shrunk transfer 7.12%).
- H162 (pinning the board's principal point hurts, re-solved). Median |e<sub>p,c</sub>| ≥ the free median |e| (3i first order 6.76%).
- H163 (pinning all four does not rescue it). Median |e<sub>p,ck</sub>| ≥ the free median |e| (3i first order 7.17%).
- H164 (first order points the right way on TUM). For each of k, c and ck, sign(f<sub>c</sub> − f) = sign(f<sub>p</sub> − f) in ≥ 90% of windows. Refuted if any of the three fails.
- H165 (first order is the right size on TUM, within 3h's overshoot). For each of k, c and ck, the median of (f<sub>p</sub> − f)/(f<sub>c</sub> − f) lies in [0.67, 1.5]. Refuted if any fails.
- H166 (the narrowed bar is real on TUM). For k, median |(sd<sub>p</sub>/sd<sub>free</sub>) − r<sub>c</sub>|/r<sub>c</sub> ≤ 0.2 (r<sub>c</sub> the first-order sd ratio), as 3h's H152 on board windows.
- Reported, unscored. Per sequence: median |e| free and pinned, median step and ratio; the robust cost rise when pinned (how hard the data resist each pin); the ratio by first-order step bin (3h's bins);
tumpin.png. - Caveat set in advance. 57 windows, three sequences, two Kinect units; the TUM reference is the board calibration the error is scored against; a pinned solve is started at the free solution and can stop in a nearby local minimum; the rebuild relies on OpenCV reproducing 3c's SIFT tracks (asserted, not assumed).
2026-11-17 · Milestone 3i result: the principal point is not what a natural-feature window's focal length compensates: pinning the board's (c<sub>x</sub>, c<sub>y</sub>) raises the TUM median error from 5.25% to 6.76% (first order), while on clean board windows the same pin, re-solved, lowers it from 0.55% to 0.37%
- Run.
tum.py (seconds; analysis only) and run.py (about 2 min on 4 cores) then analyse.py, exactly as planned below. Part A: 57 windows; the script reproduced milestone 3g's f<sub>c2</sub> for every one and the board reference reproduced milestone 4n's f and d. Part B: 149 of 150 windows converged (the same left 4-view draw as milestone 3h hit the iteration cap); every free solution equalled 3h's (asserted); all 447 principal-point re-solves and all 149 four-term re-solves converged with their pins held. tests/test_m3i.py (4 tests) passed before the run. - H154 refuted. Median |e<sub>c</sub>| 6.76% against 5.25% as solved (ratio 1.29; bar 0.8). Per sequence: fr1/desk 2.80% to 4.58%, fr1/xyz 5.18% to 7.72%, fr3/stf 7.24% to 7.26%.
- H155 refuted. Pinning (c<sub>x</sub>, c<sub>y</sub>, k<sub>1</sub>, k<sub>2</sub>) together: median |e<sub>ck</sub>| 7.17%, worse than none (5.25%), better than k<sub>1</sub>, k<sub>2</sub> alone (milestone 3g, 8.72%). fr1/xyz 12.0% (3g 16.5%).
- H156 refuted. Spearman of the signed principal-point step against the signed as-solved error −0.22 (bar ≤ −0.4): the step barely knows which way the window is wrong.
- H157 held. Median |step| 2.8% of f (four-term 6.0%), inside the range 3h checked.
- H158 held. On the board windows the first-order principal-point step matches the re-solve: median relative gap 0.089, sign agreed in 96% of windows; re-solved/first-order ratio 1.01, 1.02, 1.04 at t = 0.5, 1, 2 (steps a median 0.1–0.4% of f).
- H159 held. On clean board corners the board's principal point helps f: median |f<sub>p</sub> − f<sub>ref</sub>|/f<sub>ref</sub> 0.55% free, 0.37% pinned (ratio 0.66; better or equal in five of six camera-size cells; right camera 6 views 0.52% to 0.53%).
- H160 refuted. All four of c<sub>x</sub>, c<sub>y</sub>, k<sub>1</sub>, k<sub>2</sub> pinned: 0.47% (ratio 0.85; bar 0.6): adding the board's distortion to its principal point gives back part of the gain, as 3h's k-only pin (0.65%) suggested it would.
- Unscored. TUM windows sit far from the board's principal point: median c offset (window − board) fr1/desk (+4.1, −5.0) px, fr1/xyz (+4.9, +32.9) px, fr3/stf (+23.4, −10.5) px; on the board windows median |offsets| 1.4–7.0 px. Yet f and c are only weakly coupled in each window's covariance: median corr(f, c<sub>x</sub>) about 0.0, corr(f, c<sub>y</sub>) −0.13 to −0.28; conditional-to-marginal sd ratio 0.95–0.97. TUM's published principal point instead of the board's: median |e| 4.24%, 7.89%, 7.18% per sequence, about the same as the board's. Random-effects K = 5 pooling of the c-conditioned f: median error 2.06% (raw 2.45%) but rms z 1.27 (raw 1.10), fr1/xyz rms z 1.78.
- Reading. Two findings that look opposed and are not. On a planar board seen in 4–8 views, the window's principal point is the weak link and the board's value is a good prior (one third off the focal error), and the linearisation is accurate there. In 1.2 s of natural-feature video the windows' principal points wander tens of pixels from the board's, yet f barely leans on them (correlation −0.1 to −0.3), so pinning c moves f by a few percent in a direction unrelated to its error and adds scatter. Together with milestones 3f–3h this rules out the static intrinsics as what f compensates in short natural-feature windows: neither the distortion nor the principal point, nor both, pinned at the board's values, brings the median error under the as-solved 5.25%. What is left is outside the pinhole-plus-radial model: the Kinect's rolling shutter, the window's motion (milestone 3e found the axis swing does not predict it), or feature-track bias. For a robot the practical rule from 3f–3h stands and extends: in short natural-feature windows leave every intrinsic free and pool windows (milestone 3d); a board prior on c or k is only worth pinning when the data are board-like.
- Caveat. As planned: Part A is first order (Part B shows the principal-point linearisation is accurate at 0.1–0.4% steps on board windows; TUM's steps are about 2.8%, untested by re-solve); the TUM reference is the board calibration the error is scored against; 57 windows, three sequences; Part B is two cameras of one rig and a planar target. The TUM re-solve with c, k or both pinned is milestone 3j and still needs the video.
2026-11-17 · Milestone 3i planned: is the principal point what a short window's focal length compensates? Pin it at the board's value, first order on the TUM windows and re-solved on the board windows
Written before experiments/m3i_principal/tum.py and run.py were first run and before any principal-point-conditioned or principal-point-pinned f was computed. Milestones 3f–3h ruled out the distortion: pinning the board's k<sub>1</sub>, k<sub>2</sub> moves a natural-feature window's f the wrong way, and 3h showed that is not a linearisation artefact. The README's next candidate for what f compensates is the principal point. The TUM video is still refused by the proxy (cvg.cit.tum.de, vision.in.tum.de, cdn2.vision.in.tum.de, huggingface.co; pick_track.py host recheck today), so the TUM re-solve is renamed milestone 3j and this milestone asks the question two ways that need no new download. Seen before this plan: the board references only (Zhang k<sub>1</sub> + k<sub>2</sub> re-run on milestone 4j's stored TUM corners with milestone 4n's flags, reproducing its f to 1e-9: fr1 c = (315.7, 260.1), fr3 c = (319.5, 245.8) px).
- Part A, TUM, first order (
tum.py, analysis only). Milestone 3g's 57 scored windows and formula, f<sub>c</sub> = f + Σ<sub>f,k</sub>Σ<sub>k,k</sub><sup>−1</sup>(k<sub>ref</sub> − k), from milestone 3b's stored 5×5 covariances; the script asserts it reproduces 3g's two-term f<sub>c2</sub> for every window. e<sub>c</sub>: (c<sub>x</sub>, c<sub>y</sub>) pinned at the board's principal point; e<sub>ck</sub>: (c<sub>x</sub>, c<sub>y</sub>, k<sub>1</sub>, k<sub>2</sub>) pinned at the board's values. - H154 (the principal point is what f compensates). Median |e<sub>c</sub>| ≤ 0.8× the as-solved median |e| (5.25%). Refuted above.
- H155 (pinning everything the board knows does no harm). Median |e<sub>ck</sub>| ≤ the as-solved median |e|. Refuted above (milestone 3g's k-only pin was 8.72%).
- H156 (the principal-point step corrects the error it should). Spearman of the signed step (f<sub>c</sub> − f)/f against the signed as-solved error e ≤ −0.4 over the 57 windows (windows whose f is too high are pulled down).
- H157 (in range). Median |step| ≤ 10% of f; if refuted, H154–H156 are indicative only.
- Part B, board windows, re-solved (
run.py, analyse.py). Milestone 3h's 150 windows of OpenCV's stereo set (same seeds, subsets, starts and free solver, imported unchanged; analyse.py asserts the free solutions equal 3h's), with c<sub>t</sub> = c + t(c<sub>ref</sub> − c), t ∈ {0.5, 1, 2}, c<sub>ref</sub> Zhang's k<sub>1</sub> + k<sub>2</sub> principal point of that camera on all 13 views; f<sub>c</sub> as above and f<sub>p</sub> the re-solve with c pinned (milestone 3h's bundle_pinned); at t = 1 also the re-solve with all four of c<sub>x</sub>, c<sub>y</sub>, k<sub>1</sub>, k<sub>2</sub> pinned. tests/test_m3i.py (4 tests: pinned c stays and the true c recovers f on a synthetic scene; a 1.5-sd principal-point step re-solves within 10% of first order (it is about 5% long there, against under 5% for a 0.5-sd distortion step in 3h); the four-term conditioning against a 400,000-draw Gaussian; the board reference reproduces milestone 4n) passed before this plan. - H158 (first order predicts the principal-point re-solve). At t = 1, median |f<sub>p</sub> − f<sub>c</sub>|/|f<sub>p</sub> − f| ≤ 0.25 and sign agreement ≥ 90%.
- H159 (on clean corners the board's principal point does not hurt f). At t = 1, median |f<sub>p</sub> − f<sub>ref</sub>|/f<sub>ref</sub> ≤ the free solve's (3h: 0.55%).
- H160 (all four pinned recover the board's f). Median |f<sub>p4</sub> − f<sub>ref</sub>|/f<sub>ref</sub> ≤ 0.6× the free solve's.
- Reported, unscored. Per sequence (A) and per camera and window size (B): medians of the step, the principal-point offset from the board in px, corr(f, c<sub>x</sub>) and corr(f, c<sub>y</sub>), the conditional-to-marginal sd ratio; pinning c at TUM's published principal point (fr1 (318.6, 255.3), fr3 (320.1, 247.6)); milestone 3d's K = 5 random-effects pooling on the principal-point-conditioned f;
principal.png. - Caveat set in advance. Part A is first order under the solver's covariance shape (3h found it about a quarter long at 5–10% steps for k); part B is a planar board, not natural features; the TUM reference is the same board calibration the error is scored against. A null here narrows the list (distortion and principal point out) and leaves the rolling shutter and the window's motion as candidates.
2026-11-17 · Milestone 3h result: the first-order conditioning tracks the pinned re-solve (re-solve/first-order step 0.97 at the board's distortion, sign agreed in 99% of windows) but overshoots as the step grows (0.85 at 2–5% of f, 0.74 at 5–10%); the narrowed bar is real; shrunk by that overshoot, milestone 3g's verdict stands (median error 7.1% against 5.25% unpinned)
- Run.
experiments/m3h_linear/run.py (about 2 min on 4 cores) and analyse.py (summary.json, linear.png), exactly as planned below. 149 of 150 windows converged (left camera, 4 views, draw 4 hit the iteration cap and is excluded with its 7 pinned solves); every pinned solve of the 149 converged with k<sub>1</sub>, k<sub>2</sub> held exactly, 1043 (window, t) pairs. tests/test_m3h.py (4 tests) passed before the run. - H149 held. At t = 1, median |f<sub>p</sub> − f<sub>c</sub>| / |f<sub>p</sub> − f| 0.13; sign agreed in 98.7% of windows. But the steps at t = 1 are small here (median 0.21% of f): on these clean board windows the solved k is already near Zhang's.
- H150 held. Spearman of the gap |f<sub>p</sub> − f<sub>c</sub>|/f against the first-order step +0.69 over 1043 pairs.
- H151 held, on the minimum count. Exactly 20 pairs had a first-order step of 5–15%; median (f<sub>p</sub> − f)/(f<sub>c</sub> − f) 0.77 (bar 0.67–1.5). All 20 are steps that lower f; milestone 3g's steps raise it.
- H152 held. At t = 1, the re-solved sd of f with k pinned over the free sd (median 0.976) matches the first-order ratio (0.962) to a median 1.7%.
- H153 refuted. Pinning the board's (k<sub>1</sub>, k<sub>2</sub>) does not help even on clean corners: median |f − f<sub>ref</sub>|/f<sub>ref</sub> 0.55% free, 0.65% pinned (ratio 1.17; bar 0.6). Per camera and window size the pinned error is higher in five of six cells (left, 6 views: 0.47% to 0.45%).
- Unscored. Median re-solve/first-order ratio by first-order step: <0.25%: 0.96; 0.25–0.5%: 0.96; 0.5–1%: 0.94; 1–2%: 0.92; 2–5%: 0.85 (n 73); 5–10%: 0.74 (n 18); 10–15%: 1.03 (n 2); the overshoot is systematic, the re-solved step shorter than the predicted one. By t: 1.00, 0.99, 0.97, 0.95, 0.91, 0.88, 0.84 at t = 0.25 to 4.
- Post hoc (
post_hoc_3g in analyse.py, added after H149–H153 were scored): scaling each of milestone 3g's 57 two-term steps by this milestone's median ratio for its step bin gives median |error| 7.12% (3g first order 8.72%, unpinned 5.25%); fr1/desk 5.72%, fr1/xyz 11.39% (signed +10.4%), fr3/stf 6.19% (unpinned 7.24%). A transfer from board windows to TUM windows, indicative only. - Reading. On this solver and real images, milestone 3g's first-order formula is a fair guide to the re-solve: right in direction almost always, right in size to within a few percent below a 1% step, and about a quarter too long at 5–10%. Its narrowed error bar is the re-solve's bar too, so 3g's warning about a confidently wrong estimate is not an artefact of the linearisation. Shrinking 3g's steps by the overshoot measured here does not rescue the board prior on the Kinect (fr1/xyz still +10%), so the open question shifts from "does the linearisation mislead?" to "what does f compensate for that the window model lacks?". And even on clean board windows, where solved and board distortion nearly agree, pinning the board's distortion slightly worsens f: the free solve's own k is the better partner for its f.
- Caveat. As planned: board corners on a planar target, not natural features; two cameras of one rig; the 5–15% band holds only 20 pairs, all of one sign; the reference shares images with the windows. The TUM re-solve is now milestone 3i.
2026-11-17 · Milestone 3h planned: does the first-order conditioning of milestones 3f and 3g predict the re-solve? Checked on real board images, where the re-solve can be run
Written before experiments/m3h_linear/run.py was first run and before any pinned re-solve or conditioned f was computed. Milestone 3g's caveat was that its two-term step leaves the linearisation's range on fr1/xyz (a 13.6% step), and its re-solve needs the TUM video, which this session's proxy still refuses (CONNECT 403 on cvg.cit.tum.de, vision.in.tum.de, cdn2.vision.in.tum.de and huggingface.co, probed today). The question underneath 3f and 3g can be asked where the re-solve is possible: does Gaussian conditioning on (k<sub>1</sub>, k<sub>2</sub>) under the solver's covariance land where the pinned bundle adjustment lands, and how does the gap grow with the step? Seen before this plan: one smoke solve of the left camera's first six views (k free, f 537.8 against Zhang's 536.3), to time the solver; no conditioning or pinned solve.
- Method (
experiments/m3h_linear/). OpenCV's stereo set (data/, two cameras, 13 views each). Windows: 25 seeded random subsets of 4, 6 and 8 views per camera (150 windows). Correspondences are the detected 9×6 corners (selfcal.detect, subpix11), one track per corner, board geometry withheld. Free solve: milestone 3b's nfba.bundle unchanged (soft-L1, 1 px; f, c<sub>x</sub>, c<sub>y</sub>, k<sub>1</sub>, k<sub>2</sub>), started from selfcal.init_fronto at f<sub>0</sub> = 0.8, 1.0 and 1.25× the image width, lowest cost kept; covariance nfba.intr_cov. Reference k<sub>ref</sub>, f<sub>ref</sub>: Zhang k<sub>1</sub> + k<sub>2</sub> on all 13 views of that camera. Targets k<sub>t</sub> = k + t(k<sub>ref</sub> − k), t ∈ {0.25, 0.5, 1, 1.5, 2, 3, 4} (t = 1 is the board's distortion; larger t reaches larger steps). Per target: f<sub>c</sub>, milestone 3g's conditioned f; f<sub>p</sub>, the re-solve with k<sub>1</sub>, k<sub>2</sub> pinned at k<sub>t</sub> by the same LM with those columns dropped (pinned.bundle_pinned, started at the free solution), and its sd of f. tests/test_m3h.py (4 tests: no pins reproduces nfba.bundle; pinned values stay and the true k recovers f on a synthetic scene; a tiny step matches the conditioning to 5% and the pinned sd matches the conditional sd ratio; the one-term formula) passed before this plan. - H149 (first order predicts the re-solve at the board's distortion). At t = 1: median |f<sub>p</sub> − f<sub>c</sub>| / |f<sub>p</sub> − f| ≤ 0.25 and sign(f<sub>c</sub> − f) = sign(f<sub>p</sub> − f) in ≥ 90% of windows. Refuted if either fails.
- H150 (the gap grows with the step). Spearman of |f<sub>p</sub> − f<sub>c</sub>|/f against the first-order step |f<sub>c</sub> − f|/f over all (window, t) pairs ≥ +0.5.
- H151 (in milestone 3g's range). Over pairs whose first-order step is 5–15% of f, the median of (f<sub>p</sub> − f)/(f<sub>c</sub> − f) lies in [0.67, 1.5]. Unscored (reported as such) if fewer than 20 pairs fall in that range.
- H152 (the narrowed bar is real). At t = 1, median |(sd<sub>p</sub>/sd<sub>free</sub>) − r<sub>c</sub>| / r<sub>c</sub> ≤ 0.2, where r<sub>c</sub> is the first-order conditional-to-marginal sd ratio.
- H153 (on clean corners the board's distortion helps). At t = 1, median |f<sub>p</sub> − f<sub>ref</sub>|/f<sub>ref</sub> ≤ 0.6× the free solve's median |f − f<sub>ref</sub>|/f<sub>ref</sub>.
- Reported, unscored. Per camera and window size: medians of the step, the ratio (f<sub>p</sub> − f)/(f<sub>c</sub> − f) by t, the error before and after pinning; the binned ratio against step;
linear.png. - Caveat set in advance. Board corners on a planar target are not natural features in a 3-D scene: the windows are cleaner, the points coplanar and the covariance shape different from TUM's, so this tests the linearisation of this solver on real images, not the TUM windows themselves; the re-solve on TUM (now milestone 3i) remains the direct test. The reference shares images with the windows (Zhang uses the board geometry the windows do not get). Two cameras of one stereo rig.
2026-11-17 · Milestone 3g result: pinning both k<sub>1</sub> and k<sub>2</sub> at the board's values makes the focal error worse, not better (median 5.25% to 8.72%; fr1/xyz +12.9%), and narrows the bar by about 40% while doing it; at first order a board distortion prior does not rescue a 1.2 s window
- Run.
experiments/m3g_condk12/analyse.py (seconds; analysis only), exactly as planned below; tests/test_m3g.py (4 tests: two-term conditioning against Monte Carlo draws of a 5-D Gaussian, the one-term case equal to milestone 3f's bivariate formula, the covariance scale cancelling, uncorrelated terms leaving f) passed before the run. The script asserts that milestone 3b's lowest-cost start equals milestone 3c's solve for every window (intrinsics and sd) and that its k<sub>1</sub>-only conditioning reproduces milestone 3f's f<sub>c</sub> to 1e-6; all 57 passed. Output summary.json, condk12.png. - H144 refuted. Median |e<sub>c2</sub>| 8.72% against 5.25% as solved (ratio 1.66; bar 0.6).
- H145 refuted. Median signed e<sub>c2</sub> on fr1/xyz +12.9% (as solved −1.4%, k<sub>1</sub> alone +12.4%).
- H146 refuted. Median |e<sub>c2</sub>| against milestone 3f's k<sub>1</sub>-only |e<sub>c</sub>|: fr1/desk 8.27% against 5.75%, fr1/xyz 16.45% against 12.39%. Two terms are worse than one on both Kinect sequences.
- H147 refuted, narrowly. fr3/stf median |e<sub>c2</sub>| 7.75% against 7.24% as solved (k<sub>1</sub> alone had helped: 5.26%).
- H148 held. Median step |f<sub>c2</sub> − f|/f 6.65% over all windows (fr1/desk 5.6%, fr1/xyz 13.6%, fr3/stf 5.9%); on fr1/xyz alone the step is past the 10% bar, so that sequence's numbers are an extrapolation by this measure.
- Unscored. Median signed e<sub>c2</sub>: fr1/desk +3.3%, fr3/stf +7.6%: the two-term step pushes f up on every sequence. corr(k<sub>1</sub>, k<sub>2</sub>) in the solver's covariance, median −0.93, −0.91, −0.74 (fr1/desk, fr1/xyz, fr3/stf). Conditional-to-marginal sd of f: 0.66, 0.59, 0.57 with both terms pinned (0.69, 0.64, 0.78 with k<sub>1</sub> alone). Conditioning on k<sub>2</sub> alone: median |e| 13.7% (fr1/xyz 31%). Random-effects K = 5 blocks on f<sub>c2</sub> with sd<sub>JK</sub> times the sd ratio (assumed): all blocks rms z 3.07, median |error| 7.00% (as solved 1.10 and 2.45%); fr1/xyz 17.2%, rms z 4.05.
- Reading. Milestone 3f's rule, "pin the whole distortion model, never one coefficient", does not survive at first order: pinning the whole model is worse than pinning k<sub>1</sub>, which was already worse than pinning nothing. Within each window f, k<sub>1</sub> and k<sub>2</sub> lie along one ridge (k<sub>1</sub>–k<sub>2</sub> correlation near −0.9 on the Kinect), the board's (k<sub>1</sub>, k<sub>2</sub>) sits well off it, and the solver's covariance says getting there costs a 6–14% rise in f. Either the window's covariance shape is a poor guide that far from its solution (fr1/xyz's step is past the pre-set range), or the natural-feature windows and the board calibration describe the lens differently in a way that f compensates, for example through the principal point or the rolling shutter, which neither model nor this analysis separates. For a robot the practical reading is a warning: plugging the last board calibration's distortion into a short-window self-calibration as a prior would move f the wrong way and shrink its error bar by about 40% at once, a confidently wrong estimate (pooled rms z 3.07). Until the re-solve (milestone 3h) says otherwise, leave the distortion free in short windows and pool, as milestone 3d does.
- Caveat. As planned: first order, under the solver's covariance shape (its scale cancels, but milestone 3b found it 9× too small, and the shape may be off too); the conditional sd is assumed; 57 windows, three sequences. The two-term step is larger than 3f's one-term one, so this milestone leans harder on the linearisation; the re-solve with both terms pinned remains the test and is now milestone 3h.
2026-11-17 · Milestone 3g planned: pin both k<sub>1</sub> and k<sub>2</sub> at the board's values, first order, from the full covariances milestone 3b already stored
Written before experiments/m3g_condk12/analyse.py was first run and before any window's two-term conditioned f was computed, after milestone 3f's results were read. Milestone 3f could condition only on k<sub>1</sub> because it believed corr(k<sub>1</sub>, k<sub>2</sub>) was not stored. It is: milestone 3b's results.json keeps the full 5×5 intrinsics covariance of every start, and for all 57 scored windows the lowest-cost start reproduces milestone 3c's intrinsics and solver sd exactly (checked before this plan; largest relative difference 0). So the two-term question milestone 3f left for the re-solve can be asked to first order now. The re-solve itself (needs the TUM video; the host still refuses the proxy here, CONNECT 403 on cvg.cit.tum.de, vision.in.tum.de and cdn2.vision.in.tum.de, and huggingface.co is refused too) is renamed milestone 3h.
- Method (
experiments/m3g_condk12/). Scored set, errors e, sd<sub>JK</sub> and the board reference (milestone 4n's k<sub>1</sub> + k<sub>2</sub> fit per unit) exactly as milestone 3f. Two-term conditioned estimate (Gaussian conditioning on k = (k<sub>1</sub>, k<sub>2</sub>) = k<sub>ref</sub> under the solver's covariance Σ, c<sub>x</sub>, c<sub>y</sub> free): f<sub>c2</sub> = f + Σ<sub>f,k</sub> Σ<sub>k,k</sub><sup>−1</sup> (k<sub>ref</sub> − k). The scale of Σ cancels. Check: the same formula with k<sub>1</sub> alone must reproduce milestone 3f's f<sub>c</sub> for every window (to 1e-6 relative), or the script stops. - H144 (both terms shrink the error). Median |e<sub>c2</sub>| over the 57 windows ≤ 0.6× median |e| (milestone 3e: 5.25%, so ≤ 3.15%). Refuted above.
- H145 (fr1/xyz's offset goes). Median signed e<sub>c2</sub> on fr1/xyz within ±1.5%. Refuted outside.
- H146 (two terms beat one on the Kinect). Median |e<sub>c2</sub>| below milestone 3f's median |e<sub>c</sub>| on each of fr1/desk (5.75%) and fr1/xyz (12.39%). Refuted if either is not.
- H147 (no harm on the near-undistorted unit). Median |e<sub>c2</sub>| on fr3/stf ≤ its as-solved 7.24%. Refuted above.
- H148 (the linearisation is in range). Median |f<sub>c2</sub> − f|/f ≤ 10%; if refuted, H144–H147 are read as indicative only.
- Reported, unscored. Per sequence: medians of e<sub>c2</sub>, |e<sub>c2</sub>|, the step, corr(k<sub>1</sub>, k<sub>2</sub>) and the conditional-to-marginal sd ratio of f, √(Σ<sub>ff</sub> − Σ<sub>f,k</sub>Σ<sub>k,k</sub><sup>−1</sup>Σ<sub>k,f</sub>)/√Σ<sub>ff</sub>; conditioning on k<sub>2</sub> alone (for comparison); milestone 3d's DerSimonian–Laird K = 5 blocks on f<sub>c2</sub> with sd<sub>JK</sub> times that ratio (an assumed conditional sd, flagged as such);
condk12.png. - Caveat set in advance. First order, under the solver's covariance shape (milestone 3b found its scale 9× too small; only the shape enters f<sub>c2</sub>), and a 1.2 s window's k can sit far from the board's, which is why H148 is scored and the re-solve (3h) stays the test. The reference is the same board calibration the error is scored against. 57 windows, three sequences, a shared scene and session within each.
2026-11-16 · Milestone 3f result: a window's k<sub>1</sub> does track its focal error (Spearman +0.59), but pinning k<sub>1</sub> alone at the board's value makes fr1 worse (fr1/xyz +12.4%) and helps only the nearly undistorted fr3 (7.2% to 5.3%); the windows agree with the board on the distortion curve, not on how k<sub>1</sub> and k<sub>2</sub> share it
- Run.
experiments/m3f_condk1/analyse.py (seconds; analysis only), exactly as planned below; tests/test_m3f.py (4 tests: the conditioning formula against Monte Carlo draws of a bivariate Gaussian, the covariance scale cancelling, zero correlation leaving f unchanged, the least-squares helper) passed before the run. Every window's f matched milestone 3c's and 3e's (the script asserts it). Output summary.json, condk1.png. - H140 held. Spearman of Δk<sub>1</sub> against signed error over 57 windows +0.59 (per sequence 0.59, 0.52, 0.81).
- H141 refuted. Median |error| 5.25% as solved, 7.01% conditioned (1.34×; bar 0.6×). Per sequence: fr1/desk 2.80% to 5.75%, fr1/xyz 5.18% to 12.39%, fr3/stf 7.24% to 5.26%.
- H142 refuted. Median signed conditioned error on fr1/xyz +12.4% (as solved −1.4%).
- H143 held. The conditioning step is a median 4.4% of f (fr1/xyz 9.7%, just inside the bar), so the refutations are not an extrapolation artefact by this measure.
- Unscored. Δk<sub>2</sub> against error: Spearman −0.28; Δk<sub>1</sub> against Δk<sub>2</sub>: −0.82. Median window k<sub>1</sub>, k<sub>2</sub>: fr1/desk 0.136, −0.224; fr1/xyz 0.102, −0.155 (board 0.223, −0.568); fr3/stf 0.002, 0.008 (board −0.001, 0.075). Least squares of error on Δk<sub>1</sub> per unit: fr1 slope 0.82, intercept +8.9%, R<sup>2</sup> 0.55; fr3 slope 1.02, intercept +2.0%, R<sup>2</sup> 0.66 (with Δk<sub>2</sub> 0.55 and 0.88). Random-effects K = 5 blocks (milestone 3d's pool, conditional sd sd<sub>JK</sub>√(1 − ρ<sup>2</sup>), an assumption): all blocks rms z 3.04, median |error| 4.10% (as solved 1.10, 2.45%); fr3/stf 1.43% with rms z 0.67 (as solved 2.36%, 0.84); fr1/xyz 10.25%, rms z 4.95.
- Post hoc (added to
analyse.py after the hypotheses above were read): the radial factor 1 + k<sub>1</sub>r<sup>2</sup> + k<sub>2</sub>r<sup>4</sup> of the windows minus the board's, median per sequence, at normalised radius r = 0.2, 0.4, 0.6, 0.77 (0.77 is the image corner): fr1/desk −0.3%, −0.5%, +1.5%, +6.7%; fr1/xyz −0.4%, −0.9%, +1.1%, +7.6%; fr3/stf 0.0%, −0.1%, −0.8%, −2.4%. - Reading. Distortion and focal length are coupled within a window, as milestone 3e's covariance said, but the board's k<sub>1</sub> is not the value the window is missing. On the Kinect of fr1 the windows and the board draw nearly the same distortion curve over the inner image (within 1% out to r = 0.4) and split it differently between the two terms (window k<sub>1</sub> about half the board's, k<sub>2</sub> about a third), so moving k<sub>1</sub> alone to the board's value with k<sub>2</sub> left where the window put it bends the curve the wrong way and drags f up 10–12%; the fr1 regression puts a window with the board's k<sub>1</sub> at +8.9%. On fr3, where both fits find almost no distortion, conditioning helps (pooled K = 5 error 2.36% to 1.43%). For a robot the rule is: if a prior calibration is used to steady self-calibration, pin the whole distortion model (or the curve), never one coefficient of it. Whether pinning both k<sub>1</sub> and k<sub>2</sub> removes fr1/xyz's offset is still open and is the re-solve's question (milestone 3g, needs the TUM video).
- Caveat. As planned: first order, under the solver's covariance shape; corr(k<sub>1</sub>, k<sub>2</sub>) was not stored, so the two-term conditioning could not be done here; the conditional sd is assumed; the radial comparison is post hoc. 57 windows, three sequences.
2026-11-16 · Milestone 3f planned: if a robot conditions each window's focal length on the k<sub>1</sub> it already knows, does the error go? A first-order answer from the stored solves
Written before experiments/m3f_condk1/analyse.py was first run and before any window's k<sub>1</sub> or k<sub>2</sub> was tabulated against its error, after milestone 3e's results were read. Milestone 3e's reading pointed at distortion: |corr(f, k<sub>1</sub>)| in the solver's covariance is a median 0.63–0.77, and fr1/xyz keeps a shared offset no motion measure explains. The direct test (re-solve every window with k<sub>1</sub>, k<sub>2</sub> held at the board calibration) needs the TUM video, and the TUM host refused this session's proxy (curl: CONNECT 403, probed three times). So this milestone takes the first-order version from what is stored: milestone 3c's per-window solves (m3c_jackknife/results.json: f, k<sub>1</sub>, k<sub>2</sub> and their solver sd) and milestone 3e's corr(f, k<sub>1</sub>) from the same solves. One value was seen before this plan: fr1/desk window 0 (k<sub>1</sub> 0.052 against the board's 0.223, f 7.7% low), printed while finding the file layout; nothing else.
- Method (
experiments/m3f_condk1/). Scored set: milestone 3c's 57 windows, its errors e = (f − f<sub>ref</sub>)/f<sub>ref</sub> against milestone 4n's k<sub>1</sub> + k<sub>2</sub> board calibration per unit, and its sd<sub>JK</sub>. Reference distortion: the same board fit (fr1 k<sub>1</sub> 0.2226, k<sub>2</sub> −0.568; fr3 k<sub>1</sub> −0.0008, k<sub>2</sub> 0.0746). Conditioned estimate (Gaussian conditioning on k<sub>1</sub> = k<sub>1,ref</sub> under the solver's covariance, k<sub>2</sub> left free): f<sub>c</sub> = f + ρ<sub>f,k1</sub> (sd<sub>f</sub>/sd<sub>k1</sub>)(k<sub>1,ref</sub> − k<sub>1</sub>). The covariance's scale cancels in the coefficient, so milestone 3b's 9× under-reported sd does not enter it; only its shape does. This is a linearisation of the re-solve, not the re-solve. - H140 (k<sub>1</sub> carries the error). Spearman of Δk<sub>1</sub> = k<sub>1</sub> − k<sub>1,ref</sub> against signed e over the 57 windows ≥ +0.5. Refuted below.
- H141 (conditioning shrinks the error). Median |e<sub>c</sub>| ≤ 0.6× median |e| (milestone 3e: 5.25%). Refuted above.
- H142 (fr1/xyz's offset goes). Median signed e<sub>c</sub> on fr1/xyz within ±1.5%. Refuted outside.
- H143 (the linearisation is in range). Median |ρ<sub>f,k1</sub> (sd<sub>f</sub>/sd<sub>k1</sub>)(k<sub>1,ref</sub> − k<sub>1</sub>)|/f ≤ 10%, i.e. the correction is a modest step from the solve, not an extrapolation; refuted above, in which case H141–H142 are read as indicative only.
- Reported, unscored. Per sequence: medians of k<sub>1</sub>, k<sub>2</sub>, Δk<sub>1</sub>, e, e<sub>c</sub>; Spearman of Δk<sub>2</sub> against e; an ordinary least-squares fit of e on Δk<sub>1</sub>, Δk<sub>2</sub> per unit (descriptive, fitted on the scored data); milestone 3d's DerSimonian–Laird K = 5 blocks on f<sub>c</sub> with sd<sub>JK</sub>√(1 − ρ<sup>2</sup>) (an assumed conditional sd, flagged as such);
condk1.png. - Caveat set in advance. First order only: k<sub>1</sub> of a 1.2 s window can sit far from the board's, and the linear step may not be where the re-solve lands, which is why H143 is scored and the re-solve stays the test (milestone 3g, when the TUM host answers). corr(k<sub>1</sub>, k<sub>2</sub>) was not stored, so conditioning on both terms is not possible here. The reference k<sub>1</sub> is from the same board calibration the error is scored against, which a fielded robot would have from its last calibration. 57 windows, three sequences, shared scene and session within each.
2026-11-16 · Milestone 3e result: the optical axis swings least on fr3/stf (4.7°), not on the near-critical fr1/xyz (8.3°), and axis spread does not predict a window's focal error (Spearman +0.08); the solve reads the swing as well as motion capture does (ρ 0.98), so the signal is real but it is not the one
- Run.
experiments/m3e_axis/run.py (about 10 min on 4 cores; one re-solve per window, no jackknife) and analyse.py (summary.json, axis.png), exactly as planned below. Every one of the 57 scored windows reproduced milestone 3c's f (the script asserts agreement within 0.1%; all matched to the printed 0.01 px). tests/test_m3e.py (5 tests: roll gives zero spread, a pan gives its angle, the solver and motion-capture conventions agree, total rotation counts roll, Spearman with ties) passed before the run. - H135 refuted. Median solver axis spread D: fr1/desk 20.5°, fr1/xyz 8.3°, fr3/stf 4.7°. fr1/xyz is below half of fr1/desk but well above fr3/stf.
- H136 refuted. Spearman of D against |f error| over 57 windows +0.08 (bar ≤ −0.3). Motion-capture D gives +0.05 and total rotation +0.04.
- H137 refuted. The lower-tercile gate (D < 5.1°) drops 3 of 18 fr1/xyz windows (17%; bar ≥ 50%) and 16 of 39 fr1/desk and fr3/stf windows (41%; bar ≤ 25%), nearly all from fr3/stf.
- H138 refuted. Median |error| of kept windows 5.30% against 5.25% for all (ratio 1.01; bar 0.75).
- H139 held. Spearman of solver D against motion-capture D 0.98; medians agree to 0.4° in every sequence.
- Unscored. The axis swing is 84–94% of the total rotation (median per sequence), so roll about the optical axis is not what hid fr1/xyz in milestone 3d. D against sd<sub>JK</sub>/f: Spearman −0.27 (more swing, a somewhat tighter honest bar). |corr(f, k1)| in the solver's covariance has median 0.72, 0.77, 0.63 and |corr(f, k2)| 0.60, 0.64, 0.23 (fr1/desk, fr1/xyz, fr3/stf); neither predicts |error| (Spearman +0.07, +0.11). Milestone 3d's random-effects K = 5 blocks: all windows rms z 1.10, median-of-five error 2.29%; after the tercile gate rms z 1.37 and 2.68% (fr1/xyz 2.98%, fr3/stf 11.5% on its 5 remaining blocks). At the median gate (7.9°) fr3/stf keeps 3 of 24 windows.
- Reading. On this video, how far the camera's view direction swings in 1.2 s says nothing about how far off its self-calibrated focal length is, either within a sequence or across them: fr3/stf barely swings yet pools to within 1.2% at eight windows (milestone 3d), and fr1/xyz swings more yet stays biased. Sturm's criticality is about the motion class, and none of these hand-held windows is close to pure translation (fr1/xyz's minimum is 1.4°, median 8.3°), so the per-window geometry is not the limiting factor at 1.2 s. The error that is left looks scene- or camera-session-specific (fr1/xyz's shared offset, the near-unit f–k1 coupling on every sequence), which points at distortion, not motion: the next test fixes k1 and k2 at the camera's previous calibration, which a fielded robot does have, and asks whether fr1/xyz's offset goes. For a robot the practical rule stands from milestone 3d: pool by random effects until the sd reaches the target; a motion gate on axis swing would cost windows and buy nothing.
- Caveat. As planned: 57 windows, three sequences, windows within a sequence share a scene and session; the tercile threshold is from this data. The reading about distortion is a hypothesis from unscored correlations, not a finding.
2026-11-16 · Milestone 3e planned: does the spread of the camera's optical axes, read from the solve's own poses, mark the near-critical windows that total rotation missed?
Written before experiments/m3e_axis/run.py was first run and before any axis-spread value was computed, after milestone 3d's results were read. Milestone 3d found that total motion-capture rotation does not pick out fr1/xyz, whose windows land a shared −3% off with a jackknife sd that does not grow. Sturm (1997) lists pure translation and rotation about the optical axis among the critical motions for focal length; total rotation counts roll about the optical axis, which carries no information on f. So the candidate signal is how far the optical axis swings within the window, and a robot can read it from its own solved poses.
- Method (
experiments/m3e_axis/). Each of milestone 3c's windows is re-solved exactly as 3c did (best start from milestone 3b, OpenCV RNG seeded per window, nfba.solve_window unchanged); a window's f must reproduce 3c's within 0.1% or the script stops. Scored set: 3c's 57 scored windows, with 3c's f, error against milestone 4n's board reference and sd<sub>JK</sub>. Axis spread D = the largest angle between the optical axes of any two of the window's 12 keyframes (solver poses: x<sub>cam</sub> = R X + t, axis = third row of R), in degrees. The same D from TUM's motion capture (axis = third column of the camera-to-world rotation) is the check. Roll is excluded by construction. Also from the solve: the correlations of f with k1 and k2 in its intrinsics covariance. - H135 (D marks fr1/xyz). Median solver D on fr1/xyz ≤ 0.5× the median on each of fr1/desk and fr3/stf. Refuted otherwise.
- H136 (D predicts error). Spearman of solver D against |f − f<sub>ref</sub>|/f<sub>ref</sub> over the 57 windows ≤ −0.3. Refuted above.
- H137 (a gate on D picks the right windows). Dropping windows below the lower tercile of solver D (over all 57) removes ≥ 50% of fr1/xyz's scored windows and ≤ 25% of fr1/desk's and fr3/stf's together. Refuted if either fails.
- H138 (the gate helps). Median |error| of the kept windows ≤ 0.75× that of all 57. Refuted above.
- H139 (the solver sees what motion capture sees). Spearman of solver D against motion-capture D ≥ 0.8. Refuted below.
- Reported, unscored. Per sequence: medians of D (solver and motion capture), total rotation and the roll share of it; Spearman of D against sd<sub>JK</sub>/f and of |corr(f, k1)|, |corr(f, k2)| against |error|; milestone 3d's random-effects K = 5 blocks on the gated windows (rms z, median |error|); gate sensitivity at the lower quartile and the median;
axis.png. - Caveat set in advance. 57 windows in three sequences of one scene type, which share a scene and session within a sequence; the gate threshold is a tercile of this same data, not a fixed figure; D from 12 keyframes over 1.2 s is the swing of a hand-held camera, not designed motion.
2026-11-16 · Milestone 3d result: pool five 1.2 s windows by random effects and the sd stays honest (rms z 1.10; independence gives 2.79), but f lands a median 2.2–2.5% off; about eight windows (10 s) reach 2%, and a rotation gate does not find the near-critical case
- Run.
experiments/m3d_pool/pool.py (seconds; analysis only, on milestone 3c's summary.json), exactly as planned below; tests/test_m3d.py (5 tests: the pooling helpers on known cases, including about-unit rms z for DerSimonian–Laird under simulated random effects) passed before the run. Output summary.json, pool.png. - H130 held. Inverse-variance pooling, which treats windows as independent, gives rms z 2.79 over 45 K = 5 blocks, and it worsens with K (2.17, 2.55, 2.79, 2.90 at K = 2, 3, 5, 8).
- H131 held. DerSimonian–Laird rms z 1.10 at K = 5 (2-sd coverage 96%), and 0.89 at K = 8; per sequence 1.09, 1.39, 0.84 at K = 5.
- H132 refuted, on fr3/stf. Median |error| of the median of five windows: fr1/desk 1.48%, fr3/stf 2.29% (bar 2%), fr1/xyz 2.98% (the predicted > 2% held there).
- H133 refuted. Share of K = 5 blocks with random-effects sd ≤ 2% and |error| ≤ 2%: fr1/desk 45%, fr3/stf 0% (median sd 3.48%), fr1/xyz 0%.
- H134 refuted. The 5° gate keeps 15 of 18 fr1/xyz windows (it was to drop at least half) and drops 13 of 24 fr3/stf windows, whose gated K = 5 error rises to 6.38%. Gated rms z stays honest (1.24). At 10° it keeps 6 of 18 fr1/xyz windows and none of fr3/stf.
- Unscored. Median random-effects sd and |error| by K over all blocks: K = 1 6.45% and 5.25%; 2: 4.91%, 4.67%; 3: 4.14%, 3.76%; 5: 3.31%, 2.45%; 8: 2.53%, 1.42%. At K = 8 the median of windows is 0.25%, 2.83% and 1.18% off on fr1/desk, fr1/xyz, fr3/stf; fr1/xyz is the one sequence still above 2%.
- Reading. The between-window spread is real error, not noise to average away: a robot that treats its windows as independent will claim 2–3× too much precision, and the claim grows with every window it adds. A random-effects pool keeps the bar honest at every K tried, so the useful rule is "pool by random effects, accept when its sd reaches the target", and on this data that takes about eight windows (10 s of hand-held video) for a 2.5% sd, not one. Rotation from motion capture is not the near-criticality signal: fr1/xyz rotates as much as the others by this measure, and a rotation threshold mostly removes fr3/stf's slow, good windows. What makes fr1/xyz near-critical has to be read from the solve itself (e.g. the spread of viewing directions, which milestone 1 found predicted error on the board).
- Caveat. Sliding blocks overlap and share a scene, so the 45 blocks are far from 45 independent tests; fr1/xyz's shared offset (about −3%) is visible only against the board reference, which a robot does not have.
2026-11-16 · Milestone 3d planned: pool natural-feature windows by the keyframe-jackknife sd; how many 1.2 s windows pin f to 2%, and which pooled sd stays honest?
Written before experiments/m3d_pool/pool.py was first run on milestone 3c's data, after 3c's results were read (its summary: a single window's sd<sub>JK</sub> is honest overall but a median 6.4% of f; an inverse-variance mean over a whole sequence was −4.5% off on fr1/xyz). I had not printed the per-window rotation values before setting the gate below. Analysis only: no solve is re-run.
- Method. Inputs: 3c's 57 scored windows (
m3c_jackknife/summary.json: f, sd<sub>JK</sub>, motion-capture rotation). Within a sequence the scored windows are put in time order; a block of size K is K consecutive scored windows, sliding by one (K = 1, 2, 3, 5, 8; headline K = 5, about 6 s of video). Each block gives (i) the inverse-variance mean under sd<sub>JK</sub> with sd 1/√Σw (windows treated as independent), (ii) the DerSimonian–Laird random-effects mean with sd 1/√Σw* (adds the between-window spread τ<sup>2</sup>), (iii) the plain median (milestone 3b's pooling; point only). Reference and z as 3c (milestone 4n's board f per unit). Gate: windows with motion-capture rotation below 5° are dropped before blocking, as a stand-in for a gyro reading (3° and 10° reported unscored). - H130 (independence overclaims). Inverse-variance rms z over all K = 5 blocks > 2. Refuted at or below 2.
- H131 (random effects is honest). DerSimonian–Laird rms z over all K = 5 blocks ≤ 2. Refuted above.
- H132 (pooling helps where motion is rich, not on near-critical motion). The median of five windows has median |error| ≤ 2% on fr1/desk and fr3/stf, and > 2% on fr1/xyz. Refuted if any of the three fails.
- H133 (a usable answer in about 6 s). On fr1/desk and fr3/stf, at least half of K = 5 blocks have a random-effects sd ≤ 2% of f and |error| ≤ 2%. Refuted below half in either.
- H134 (a rotation gate removes the near-critical case). The 5° gate drops at least half of fr1/xyz's scored windows, and random-effects rms z over the gated K = 5 blocks is ≤ 2. Refuted if either fails.
- Reported, unscored. Every K: rms z, 2-sd coverage, median sd and |error| per method and sequence; gate sensitivity at 3° and 10°;
pool.png. - Caveat set in advance. Sliding blocks overlap, so blocks are not independent (fr1/desk has 15 scored windows, fr1/xyz 18, fr3/stf 24, giving 11, 14 and 20 K = 5 blocks); windows of one sequence share a scene and a session; a DerSimonian–Laird τ<sup>2</sup> from five windows is itself noisy; the gate uses motion capture, which a robot would replace with an IMU.
2026-11-16 · Milestone 3c result: a delete-one-keyframe jackknife makes a natural-feature window's sd of f honest overall (rms z 1.76, from the solver's 9.3; 2-sd coverage 89%), but the honest sd is a median 6.4% of f, so one 1.2 s window rarely pins f to 2%; per-track correlation is not the cause (track sandwich rms z 6.6)
- Runs.
experiments/m3c_jackknife/run.py (about 50 min on 4 cores, after a first attempt was stopped by the 30-minute background limit before it wrote anything; the rerun saves each window as it finishes and skips the jackknife on windows that already fail milestone 3b's counting rules, which are not scored) re-solved all 65 windows from milestone 3b's best start with a seeded RNG, then 12 warm-started delete-one-keyframe solves and the two sandwiches per counted window; analyse.py scores (summary.json, jackknife.png). One window (fr1/desk 3) was run once beforehand as a timing and plumbing pilot; only its solve statuses and iteration counts were read. tests/test_m3c.py checked first, on synthetic scenes, that re-anchoring keeps every projection, that deleting a keyframe re-indexes cleanly, that under i.i.d. noise the track sandwich matches the solver's sd within 1.5× and the jackknife within 2×, and that the lag-1 statistic reads 0.8 on injected per-track offsets and 0 on independent noise. - Reproduced. All 64 windows milestone 3b solved re-solve to within 0.1% of its f (median difference 0); the same 57 windows count. 8 of the 684 delete-one solves hit the iteration cap; no window lost its jackknife.
- H124 held. rms z<sub>JK</sub> over 57 windows 1.76 (solver 9.31, track sandwich 6.63, keyframe sandwich 3.03). Coverage at 1 and 2 sd: 60% and 89% (solver 7% and 18%).
- H125 refuted, on fr1/xyz. rms z<sub>JK</sub>: fr1/desk 1.51, fr1/xyz 2.58, fr3/stf 0.98.
- H126 held. Across-window spread over median sd<sub>JK</sub>: 0.89, 1.60, 1.19 (solver 4.2, 8.4, 11.7).
- H127 refuted. Median sd<sub>JK</sub>/f on fr1/xyz 5.74% against fr1/desk 3.73% (1.54×; bar 2×; solver 1.28×).
- H128 refuted. Track-clustered sandwich rms z 6.63 (per sequence 4.8, 8.9, 5.5). Lag-1 residual correlation along tracks, median per sequence (x, y): fr1/desk −0.12, 0.14; fr1/xyz 0.06, 0.04; fr3 −0.15, 0.06, so tracks carry little correlated error once points are refitted (the synthetic test shows a constant per-track offset is largely absorbed by the point).
- H129 refuted. Only 5 of 57 windows (9%) have sd<sub>JK</sub>/f ≤ 2%, and their median error is 2.8%.
- Unscored. Median sd<sub>JK</sub>/f: 3.7%, 5.7%, 9.1% (solver 0.85–1.09%). Keyframe-clustered sandwich: rms z 2.64, 3.98, 2.36, so per-keyframe clustering gets most of the way and the jackknife's nonlinearity the rest. Spearman of sd<sub>JK</sub> against rotation: 0.22, 0.15, 0.35 (none tested). Against TUM's calibration, rms z<sub>JK</sub> 1.74 overall. Inverse-variance mean under sd<sub>JK</sub>: fr1/desk 513.1 px (−0.99%), fr1/xyz 494.7 px (−4.5%), fr3 531.6 px (+0.16%); the fr1/xyz mean is pulled to the low side by the windows the jackknife calls tightest, so weighting by sd<sub>JK</sub> does not rescue near-critical motion, where milestone 3b's plain median (−1.4%) was better. Post hoc, chosen after scoring: the single most influential keyframe carries a median 38–46% of a window's jackknife variance (over half in 28–46% of windows), so a window's f hinges on one or two keyframes.
- Reading. H5's failure in milestone 3b was a covariance problem, and the cure is the one milestone 3 found on board corners: resample keyframes. The solver's sd assumes independent pixel noise around a linear model; the error that matters here enters per keyframe (one frame's pose, blur or rolling shutter moves the whole solve) and through nonlinearity, which only a re-solve sees. Track correlation, the first candidate, is ruled out. The honest answer is sobering for a robot: a 1.2 s hand-held window gives f to about ±4–9% (1 sd), and only 9% of windows reach 2%. So the useful output of one window is the honest interval, which tells the robot to pool windows, and the near-critical case still escapes: on fr1/xyz's translation-only motion the jackknife stays too small (rms z 2.6) and does not grow as the motion weakens (1.54×). A motion check (rotation from the IMU or the poses) remains needed for that case.
- Caveat. Same 65 windows and references as milestone 3b; windows of a sequence share a scene and are not independent. A 12-keyframe jackknife has about 11 degrees of freedom, so single-window sd<sub>JK</sub> is noisy; the hypotheses pool over windows for that reason. H124 held at 1.76 against a bar of 2, driven by fr3's 0.98.
2026-11-16 · Milestone 3c planned: on natural-feature video, does a jackknife over keyframes give each window an honest sd for f, where the solver's sd was 4–12× too small?
Written before any milestone-3c quantity was computed, after milestone 3b's results were read. Milestone 3b found that one 12-keyframe window of TUM RGB-D video gives f to within 3–7% (median) but its bundle-adjustment sd understates the error 4–12× (rms z 9.3) and does not grow on the near-critical fr1/xyz motion. On board corners, milestone 3 found the delete-one-view jackknife honest where the solver's sd was 2× small. Milestone 3b named three candidate causes: SIFT errors correlated along a track, per-frame effects (rolling shutter, blur), and weak motion the linearised sd misses. This milestone asks whether a covariance that does not trust the residual model fixes it, and which clustering does: per keyframe (delete-one-keyframe jackknife; a keyframe-clustered sandwich, unscored) or per track (a track-clustered sandwich).
- Method (
experiments/m3c_jackknife/). Every one of milestone 3b's 65 windows is re-solved by its nfba.solve_window, unchanged, from the start that had the lowest cost in milestone 3b, with OpenCV's RNG seeded per window (milestone 3b did not seed it, so RANSAC can pick slightly different tracks; agreement with milestone 3b's f is reported). A window is scored if the re-solve passes milestone 3b's counting rules. Then, warm-started from the full solution: 12 delete-one-keyframe solves (tracks keep ≥ 2 remaining views; deleting keyframe 0 re-anchors the gauge on keyframe 1), with sd<sub>JK</sub> = √((n−1)/n Σ(f<sub>i</sub> − f̄)<sup>2</sup>) over the delete-one solves that converge with 300 < f < 1000; with fewer than 10 such, sd<sub>JK</sub> is undefined and the window is left out of H124–H129 (reported). Sandwiches: S<sup>−1</sup>BS<sup>−1</sup> on the camera block after eliminating points, B summed from the points-adjusted robust scores per cluster with a G/(G−1) factor. Reference as milestone 3b: milestone 4n's board calibration (fr1 518.28, fr3 530.71 px); z = (f − f<sub>ref</sub>)/sd. - H124 (the jackknife is honest overall). rms z<sub>JK</sub> over scored windows ≤ 2 (solver: 9.3). Refuted above.
- H125 (honest in every sequence). rms z<sub>JK</sub> ≤ 2 in each of fr1/desk, fr1/xyz, fr3/stf. Refuted above in any.
- H126 (it follows the real scatter). In each sequence, the robust spread of f across windows (1.4826 MAD) over the median sd<sub>JK</sub> lies in [0.5, 2] (solver: 4.2–11.7). Refuted outside in any.
- H127 (it sees weak motion). Median sd<sub>JK</sub>/f on fr1/xyz ≥ 2× that on fr1/desk (solver 1.28×). Refuted below.
- H128 (per-track correlation is the cause). rms z under the track-clustered sandwich ≤ 2 overall. Refuted above.
- H129 (a usable gate). Windows with sd<sub>JK</sub>/f ≤ 2% are at least 25% of scored windows, and their median |f − f<sub>ref</sub>|/f<sub>ref</sub> is at most 2%. Refuted if either fails.
- Reported, unscored. Per sequence: rms z, 1- and 2-sd coverage under the solver, sd<sub>JK</sub>, the track-clustered and the keyframe-clustered sandwich; z against TUM's published calibration; Spearman of each sd against the window's motion-capture rotation; lag-1 correlation of residuals along a track between consecutive keyframes (inliers); an inverse-variance mean of each sequence's windows under sd<sub>JK</sub>; failures of delete-one solves by cause.
- Caveat set in advance. Windows of a sequence share a scene and session, so they are not independent draws; the two references differ by 0.3% (fr1) and 1.2% (fr3). A jackknife over 12 keyframes is itself noisy (about 11 degrees of freedom), so single-window sd<sub>JK</sub> varies; H124–H126 pool over windows for that reason. A keyframe-clustered sandwich with 12 clusters and per-keyframe pose parameters is biased low (each keyframe's own pose absorbs part of its cluster score), which is why it is unscored.
2026-11-16 · Milestone 4u result: centring the 5% withhold window on the unit's own 13-view calibration keeps every one of 18 cells under 1.1% false alarms (pooled 0.26%) and makes withholding symmetric (±4%: 19–22% on both units, from 11–71%), lifting ±4% detection 1.25× over rule W; ±2% power stays at 0.73× ungated
- Runs.
experiments/m4u_owncal/analyse.py (about 6 min on 4 cores, no new solve) imports milestone 4t's code unchanged, rebuilds the 18 cells for both twins, and applies rule U, rule W, the out-of-sample anchor and no withholding to the same fired masks (summary.json, owncal.png). It reproduced milestone 4t first (W pooled S0 0.30%, ungated 1.61%, W ±2% detection 3.89%), which the script asserts. tests/test_m4u.py checked first the two anchors, U's edges and its symmetry under s, and that U at f<sub>P</sub> is 4t's W. - Anchors. f<sub>U</sub> (13-view f of the old pool): fr1 A 525.16, B 526.04, C 521.09 px; fr3 A 534.04, B 537.14, C 534.46 px.
- H118 held. S0 under U, per cell: fr1 AA 0.29%, BB 0.12%, CC 0.26%, AB 0.24%, BA 0.18%, AC 1.05%, CA 0.31%, BC 0.43%, CB 0.19%; fr3 AA 0.06%, BB 0.38%, CC 0.13%, AB 0.25%, BA 0.30%, AC 0.13%, CA 0.12%, BC 0.00%, CB 0.26%. All under 2.5%.
- H119 held. Pooled 0.26% (W 0.30%, ungated 1.61%).
- H120 held. U withholds 5.35% of usable splits at s = 1 (fr1 4.0%, fr3 6.7%; W 5.8%).
- H121 held. Share withheld at +4% / −4%: fr1 19.1% / 21.6% (ratio 1.13; W 11.1% / 38.1%), fr3 19.5% / 22.2% (1.14; W 71.4% / 10.9%).
- H122 refuted. ±2% detection 3.91% against 5.34% ungated (0.73×, the same as W's 3.89%).
- H123 held, at the bar. ±4% detection 14.20% against W's 11.36% (1.250×; bar 1.25×; ungated 19.06%). By side: fr3 +4% 11.7% (W 1.4%), fr1 −4% 16.8% (W 11.1%); the near-side gains are partly offset on the far sides (fr1 +4% 15.6% against W 18.0%, fr3 −4% 12.7% against 14.9%).
- Unscored. The out-of-sample anchor (the pool in neither role) gives the same picture: pooled S0 0.26%, worst cell 1.05%, withheld 5.39%, ±2% 3.92%, ±4% 13.9%, so the in-sample overlap of f<sub>U</sub> with the old set does not drive U's figures. U agrees with W on 94–100% and with milestone 4q's oracle flag on 98–100% of s = 1 splits per cell (W: 94–100%). Window sweep under U (pooled S0, worst cell, withheld at s = 1, ±2%, ±4%): 3% 0.23%, 1.06%, 11.6%, 2.79%, 0.41%; 4% 0.24%, 1.06%, 7.3%, 3.67%, 5.7%; 5% as above; 6% 0.30%, 1.11%, 4.5%, 3.98%, 16.9%; 8% 0.32%, 1.11%, 3.7%, 4.01%, 17.7%. So on these pools a window of 6–8% stays as safe as 5% and keeps 89–93% of the ungated ±4% firing; that was read after scoring and is a candidate to pre-register, not a result. Resid-only twin under U: pooled S0 1.00%, worst cell 4.68% (fr3 BB again), so the jitter is still needed.
- Reading. Centring on the unit's own calibration fixes what milestone 4t's reading blamed on the fleet prior: withholding is now even on both sides (ratios 1.13 and 1.14), and the Xtion catches a +4% change in 11.7% of splits instead of 1.4%. It did not move ±2% power, which confirms milestone 4t's post hoc reading that the 2% loss is the far solves' spurious firing (counted as detection when every split is decided), not the window's placement; at 2% a 5% window rarely binds on a real change. The remaining ±4% loss is the window itself: a 4% change applied to a 6-view solve already 1% high leaves a 5% window, which is why wider windows recover it on this data. For a robot, U needs nothing it lacks (its own calibration and its two small-set solves) and is the better default than the spec-sheet window.
- Caveat. Same 18 cells as milestones 4p–4t, so the far-solve mechanism, rule W and now U were all named and scored on these six pools; a new session or unit remains the test. H123 held exactly at its bar (1.2504×). Splits share views; no interval.
2026-11-16 · Milestone 4u planned: if the withhold window is centred on the unit's own current calibration instead of the fleet prior, is the update test still safe in every pairing, and does it get back the power rule W lost on the side where each unit's f sits?
Written before any milestone-4u quantity was computed, after milestone 4t's results were read. Milestone 4t's rule W withheld the decision when the old 7-view f or the scaled new 6-view f lay more than 5% from the fleet prior f<sub>P</sub> = 527.1 px, and it made every cell safe (at most 1.11%) but caught ±2% changes at only 0.73× the ungated rate, because f<sub>P</sub> sits 1–2% from each unit's own f and so a 4% change toward the window's near edge was withheld most of the time (fr3 +4%: 71%, −4%: 11%). A robot also holds its own current calibration, so this milestone tests rule U: withhold when the old 7-view f or s · f<sub>new</sub> lies more than 5% from f<sub>U</sub>, the 13-view f of the old pool (the full calibration the unit was commissioned with on the views the old set is drawn from). Everything else is milestone 4t's code unchanged: the same pools, solves, own-τ jittered twin ("jit", scored), resid-only twin (unscored), 7 + 6 design, 1,716 split pairs, per-split 99th-percentile threshold, 18 cells. The run must first reproduce milestone 4t's figures (rule W pooled S0 0.30%, ungated 1.61%, W ±2% detection 3.89%). Rates are defined as in 4t (false alarms over decided splits; detection over every usable split, a withheld split counted as not detected).
- H118 (U is safe in every cell). S0 at most 2.5% in each of the 18 cells. Refuted above in any.
- H119 (near nominal). S0 pooled over the 18 cells at most 1%. Refuted above.
- H120 (withholding is rare). At s = 1, U withholds at most 10% of usable splits pooled. Refuted above.
- H121 (U withholds symmetrically). On each unit, pooled over its nine cells, the share withheld at s = 1.04 and at s = 0.96 differ by at most a factor of 2 (larger over smaller; W: fr1 3.5, fr3 6.5). Refuted above 2 on either unit.
- H122 (U keeps the power). ±2% detection pooled under U at least 0.8× the ungated jit arm's. Refuted below.
- H123 (U beats W at ±4%). ±4% detection pooled under U at least 1.25× W's (4t: 11.4%). Refuted below.
- Reported, unscored. Per cell: S0, share withheld at each s and detection at ±1, 2, 4% under U, W and ungated; agreement of U with W and with milestone 4q's oracle flag; the same under an out-of-sample anchor (the 13-view f of the unit's pool that is in neither the old nor the new role, the mean of the other two for the same-pool cells) and for windows of 3, 4, 6 and 8%; the resid-only twin under U.
- Caveat set in advance. f<sub>U</sub> is fitted on 13 views that include the old set's 7, so the old solve's distance from it is partly in-sample, which is why the out-of-sample anchor is reported. No new views: the 18 cells are milestones 4j, 4p and 4s's pools, already used to name the far-solve mechanism and to score W. The same 5% window means a 4% change still lands 1% from the edge, so some withholding of real changes remains by construction. Splits share views; no interval.
2026-11-16 · Milestone 4t result: withholding the decision when either small-set solve is more than 5% from the fleet prior brings every one of 18 cells under 1.2% false alarms (pooled 0.30%, from 1.61%) at 5.8% of splits withheld, but costs a quarter of the ±2% detections and over half the ±4% ones, because a fixed anchor withholds real changes on the side where the unit's f already sits
- Runs.
experiments/m4t_withhold/analyse.py (about 6 min on 4 cores, no new solve) rebuilds milestone 4s's pools, solves, twins and per-split thresholds for all nine pairings of A, B and C on each unit, arms "jit" (scored) and "resid" (unscored), and applies rule W to the decision only (summary.json, withhold.png). It reproduced milestone 4s first (jit fr1 CC S0 7.47%, jit pooled S0 over the ten C cells 2.60%), which the script asserts. tests/test_m4t.py checked first that f<sub>P</sub> is milestone 4k's 527.1 px and W's edges, including that the robot's scaled new f (s · f<sub>new</sub>) is what is compared. - H113 held. S0 under W, per cell: fr1 AA 0.29%, BB 0.12%, CC 0.32%, AB 0.36%, BA 0.24%, AC 1.11%, CA 0.31%, BC 0.56%, CB 0.31%; fr3 AA 0.06%, BB 0.46%, CC 0.13%, AB 0.25%, BA 0.31%, AC 0.13%, CA 0.12%, BC 0.00%, CB 0.26%. All under 2.5% (ungated, the same twin: up to 7.47%).
- H114 held. Pooled over the 18 cells: 0.30% (ungated 1.61%).
- H115 refuted, by one cell. fr1's C cells: CC 0.32%, CA 0.31%, BC 0.56%, CB 0.31% and AC 1.11% (bar 1%; ungated 5.73%).
- H116 held. W withholds 5.80% of usable splits at s = 1 (fr1 3.8%, fr3 7.8%).
- H117 refuted. ±2% detection, withheld counted as missed: 3.89% against 5.34% ungated (0.73×; bar 0.8×).
- Unscored. W and milestone 4q's oracle flag (far from the pool's median of that size) agree on 94.4–100% of splits per cell; pooled S0 under the oracle 0.27% (W 0.30%). Share withheld by scale, pooled: 24.5% at −4%, 6.2% at −2%, 5.8% at 0, 10.1% at +2%, 41.3% at +4%; by unit, fr1 withholds 38% at −4% (11% at +4%) and fr3 71% at +4% (11% at −4%): the side where each unit's f already sits from f<sub>P</sub> (fr1 −0.2% to −1.1%, fr3 +1.3% to +1.9%). Detection under W against ungated: ±1% 1.03% (2.39%), ±4% 11.4% (19.1%); at +4% on fr3 1.4% (13.7%). Resid-only twin under W: pooled S0 1.03% (ungated 2.34%), worst cell 4.84% (fr3 BB, the flat Xtion pool of milestones 4q and 4r), so W does not replace the jitter. Post hoc, computed after scoring: firing above the false-alarm floor, ±2% detection minus S0, is 3.61 points under W against 3.73 ungated (0.97×), so most of H117's loss is the far-solve firing that ungated ±2% counts as detection.
- Reading. The two failure modes milestones 4r and 4s found need two fixes, and together they work: the per-view focal jitter makes the threshold honest where the twin's spread was too narrow, and withholding the decision on a far small-set solve removes the near-degenerate view sets the linear twin cannot represent. With both, every pairing of three pools on two cameras sits at or under 1.11% false alarms, and W needs nothing a robot lacks (its two solves and the spec sheet's f), where milestone 4s's diagnostic needed every subset of a pool. Its cost is where the window is centred: f<sub>P</sub> is a fleet figure 1–2% from each unit's own f, so a 4% change toward the window's near edge is withheld most of the time (fr3 +4%: 71%). A real update of that size would be asked for more views rather than missed silently, which is the right failure for a robot, but it is a failure of power; a window centred on the unit's own current calibration should keep the symmetry, and is the next thing to pre-register.
- Caveat. No new views: the 18 cells are the pools of milestones 4j, 4p and 4s, the far-solve mechanism was named on the same data, and the oracle's rates on the C cells were read before the plan. The one cell over 1% (fr1 AC) is within 0.11 points of the bar; with splits sharing views no interval is given. The TUM host refused the proxy (HTTP 403), so a new unit or session was not tried.
2026-11-16 · Milestone 4t planned: if the update test withholds its decision whenever either small-set solve lands more than 5% from the fleet's nominal focal length, is it safe in every pairing of the three pools, and what does withholding cost in power?
Written before any milestone-4t quantity was computed, after milestone 4s's results were read. Milestone 4s found the Kinect's third pool over-alarming at up to 7.5% whatever per-view jitter its twin carried, and a post hoc diagnostic put 83–97% of those false alarms on the splits whose old 7-view or new 6-view solve is a "far" solve (more than 5% from the median solve of its size in its own pool, milestone 4q's flag). That flag is an oracle: it needs the median over every subset of a pool a robot only partly has at decision time, and its rates on the C cells were read before this plan. Milestone 4k showed the far solves are true minima of near-degenerate view sets (a warm start lands on every one of them), so they cannot be solved away, and that clipping the delete-one solves to a window around the fleet prior does harm (it narrows R). This milestone tests the deployable version on the decision only: rule W withholds the decision (no update, no alarm, ask for more views) when the old 7-view f or the scaled new 6-view f (s · f<sub>new</sub>, what the robot sees) lies more than 5% from f<sub>P</sub> = 527.1 px, milestone 4k's fleet prior (TUM's mean published f over the two units); R, the twin and the threshold are untouched. Everything else is milestone 4s's code imported unchanged: the own-τ jittered twin ("jit", milestone 4r's pool_twin), the 7 + 6 design's 1,716 split pairs, 4i's per-split 99th-percentile threshold on z<sub>R</sub><sup>2</sup>, 4p's decide, ungated. All 18 cells are scored: AA, BB, CC, AB, BA, AC, CA, BC, CB on each unit. The run must first reproduce milestone 4s's figures (jit fr1 CC S0 7.47%, jit pooled S0 over the ten C cells 2.60%).
- Rates. A false-alarm rate is the share of decided (not withheld) splits at s = 1 that fire. A detection rate counts every usable split and scores a withheld one as not detected, since the robot did not update.
- H113 (W is safe in every cell). S0 at most 2.5% in each of the 18 cells. Refuted above in any.
- H114 (near nominal). S0 pooled over the 18 cells at most 1%. Refuted above.
- H115 (W fixes the Kinect's third pool). S0 at most 1% in each of fr1's five C cells (jit, ungated: 2.9–7.5%). Refuted above in any. Partly foreseen: the oracle flag's no-far rate over those cells was 0.49%; W's flag differs (a fixed anchor, not the pool median) and is computed here for the first time.
- H116 (withholding is rare). At s = 1, W withholds at most 10% of usable splits pooled over the 18 cells. Refuted above.
- H117 (withholding keeps the power). Detection at ±2% pooled over the 18 cells, under W, is at least 0.8× the ungated jit arm's (every split decided). Refuted below 0.8×. Some loss is expected by construction, since ungated firing on far splits is counted as detection.
- Reported, unscored. Per cell: S0 under W and under the oracle flag, share withheld at each s, detection at ±1%, ±2%, ±4%; how often W and the oracle flag agree; the same for the resid-only twin; harmful adoptions (s = 1 splits where the test fires on a far solve) under W against ungated.
- Caveat set in advance. No new views: the three pools of each unit are those of milestones 4j, 4p and 4s, so the fourteen A/B cells were already the basis of milestones 4p–4r, and the far-solve mechanism was named on the same data this rule is scored on. f<sub>P</sub> sits −1.2% to +1.4% from the units' 13-view f, so W's window is asymmetric per unit (about −3.8% to +6.4% of fr1's f, −6.4% to +3.6% of fr3's); a 4% change applied to the new views can itself be withheld, which H117's detection counts. Splits share views; no interval. The TUM host refused the proxy this session (HTTP 403), so a new sequence or unit was not available.
2026-11-16 · Milestone 4s result: on a third pool of views the per-view focal jitter keeps the Xtion safe (at most 0.53% false alarms) but not the Kinect (up to 7.5%), whose third pool fits a small jitter and still over-alarms; borrowing τ does not help, and the three 13-view calibrations agree within the twin's sd
- Runs.
experiments/m4s_third/prepare.py reproduced pools A (milestone 4j) and B (milestone 4p) exactly on both units and picked C (gap 0.5 s, see the amended plan); run.py solved 9,440 k1 + k2 subsets of C (about 1 h 55 min on 4 cores; all 13-view solves converged, 5-view 1,283 and 1,282 of 1,287); analyse.py (about 10 min on 3 cores) runs milestone 4r's pool_twin, 4p's decide and 4i's per-split threshold unchanged on all three pools (summary.json, third.png). Its A/B cells reproduce milestone 4r first (jittered fr3 BB S0 0.58%, resid-only pooled ±2% 9.08%), which the script asserts. The hypotheses were planned as H104–H108 and renumbered H108–H112 at merge, unchanged in content, because milestone 3b, planned and landed in parallel, had already used H104–H107. tests/test_m4s.py checked first that C is disjoint from A and B and keeps the 0.5 s gap, the pick rule's two exclusions, that A and B load as before, and the borrowed-τ conversion. - Pool C. Median tilt fr1 0.244 (A 0.268, B 0.147), fr3 0.083 (A 0.101, B 0.063), so C sits between A and B on both units. Fitted τ (share of f): fr1 0.77% (A 0.75%, B 1.88%), fr3 0 (A 0.22%, B 1.80%): fr3 C's twin already spreads its 6-view f as widely as the real solves (1.11% against 1.10%), fr1 C's needs a small jitter (twin 0.46% against real 0.63%).
- H108 refuted, on the Kinect only. Own-τ jittered twin, S0 per C cell: fr1 CC 7.47%, AC 5.73%, CA 2.93%, BC 5.14%, CB 2.97% (all five above 2.5%); fr3 CC 0.47%, AC 0.29%, CA 0.41%, BC 0.12%, CB 0.53%.
- H109 refuted. Pooled over the ten cells: 2.60% (bar 1%); fr3's five alone are all under 0.6%.
- H110 refuted. C jittered with A's and B's mean τ (fr1 1.31%, fr3 1.01% of f): fr1 CC 7.18%, AC 5.44%, CA 2.93%, BC 5.02%, CB 3.09%; fr3 0.18–0.53%. Pooled 2.55%. A larger τ on the Kinect's C moves its false alarms by less than 0.3 points.
- H111 held. Resid-only twin: fr1 CC 7.71%, AC 6.44%, CA 3.46%, BC 5.84%, CB 3.96%; fr3 CC 0.47%, AC 0.23%, CA 0.23%, BC 1.34%, CB 1.11%. Pooled 3.08%. So the jitter is not what keeps the Xtion's C cells safe (they already were, apart from the cells holding B) and does not rescue the Kinect's.
- H112 held. 13-view f, in the jittered twins' sd: fr1 C against A −1.38, against B −0.84; fr3 +0.11 and −0.42. In share of f: fr1 C 521.09 is −0.77% from A and −0.94% from B (resid-only twin's sd: −2.10 and −1.59), fr3 C 534.46 is +0.08% and −0.50%. Against TUM's published f: fr1 C +0.81% (A +1.60%, B +1.77%), fr3 C −0.53% (A −0.61%, B −0.03%).
- Unscored. Pooled ±2% over the C cells: resid-only 8.84%, own τ 6.36% (0.72×), borrowed 5.15% (0.58×), f + principal point 7.71% (0.87×); ±4%: 26.9%, 20.1%, 16.0%, 24.0%. Milestone 4q's gate (1% cut) on the jittered set sd: 17.3% of C splits scored, S0 4.36% pooled, worst cell 6.55%, ±2% 14.7% (resid-only gate: 35.7% scored, S0 5.69%). Far solves (more than 5% from the median of their size) at 5, 6 and 7 views: fr1 A 29, 6, 1; B 72, 34, 8; C 56, 96, 62; fr3 A 46, 19, 1; B 202, 112, 38; C 197, 118, 46. Splits holding a far solve (milestone 4q's flag, old or new set; a diagnostic added after the scored figures were read, which it leaves unchanged): fr1's C cells hold 65–157 of about 1,710 each, and with the own-τ twin 48–85% of them fire against 0.26–1.05% of the rest (CC 79.0% against 0.26%; AC 85.3% against 1.05%); they carry 83–97% of every fr1 C cell's false alarms. Without them the jittered S0 over fr1's five C cells is 0.49% (resid-only 1.01%), fr3's 0.13% (0.52%). On fr3 the far splits are as many (65–162 per C cell) but fire in only 1.3–7.7% of cases, because its twin's spread is already as wide as the real one.
- Reading. The test the programme has waited on since milestone 4m came out split by camera. On the Xtion, the third pool behaves as milestone 4r's reading predicts: a flat pool whose twin already carries its spread is safe with or without the jitter, and every cell, including C against the flat pool B, stays under 0.6% once the pools are jittered. On the Kinect the per-view focal jitter does not transfer: C is a well-tilted pool whose robust 6-view spread is close to A's, so the fitted τ is small (0.77%), yet its 7-view solves hold 62 far solutions where A's hold one, and the update test fires on 3% to 7.5% of unchanged splits whatever τ the twin is given (borrowing a larger one changes it by less than 0.3 points). A robust spread that ignores the far solves cannot see this error, so the jitter, fitted to that spread, cannot fix it; the diagnostic shows it directly, since the splits holding a far solve carry nearly all of fr1 C's false alarms and the rest sit at 0.5%: the Kinect's over-alarm is the far-solve mechanism milestone 4p first named, which milestone 4q found absent on the Xtion's flat pool, back on a different pool and a different camera. For a robot, milestone 4r's one-number fix is not enough on its own: a pool's far-solve count is a second thing it has before testing, and on this evidence the threshold must account for it (or the far sets be withheld) as well as for the per-view jitter. The three 13-view calibrations of each camera agree within the jittered twin's sd (at most 1.4), so the full calibrations are consistent; the trouble is the small-set solves the test compares.
- Caveat. C comes from the same hand-held sweeps as A and B, and its frames are as close as 0.5 s to theirs; the gap was halved from 4p's 1.0 s before any C frame was picked, because 1.0 s left no candidate in 12 of 13 blocks. Splits share views; no interval. The far-solve diagnostic was added after the five hypotheses were scored, so the reading it supports is a description, not a tested hypothesis; it is the hypothesis the next milestone should pre-register.
2026-11-16 · Milestone 4s planned: does the per-view focal jitter keep the update test safe on a third pool of views that played no part in suggesting it?
Written before any milestone-4s pool was picked or solved, after milestone 4r's results were read. Milestone 4r's fitted per-view focal jitter brought every cell's false alarms under 0.6%, but on the two pools (A, B) whose over-alarm motivated it, so its caveat asked for a third disjoint pool; milestones 4m and 4q could not get one because the TUM host refused the proxy, and this session's probe got through. Pool C: a third 13 views per unit from the same two sequences (experiments/m4s_third/prepare.py), milestone 4p's detection, 13 time blocks and pick rule imported unchanged, which must reproduce A and B exactly; in each block the frames A and B kept and every frame within 1.0 s of either are removed before the pick. Amended before any C frame was picked or solved: the 13 time blocks span only 2.1 s (fr1) and 2.3 s (fr3) apart from one long block each, so 1.0 s either side of two frames left no candidate in 12 of 13 blocks on each unit (counts per block printed by a dry run of the pick, nothing else); the gap is 0.5 s, which leaves 11 to 38 candidates in every block. C's views are therefore closer in time to A's and B's (at least 0.5 s) than B's are to A's (at least 1.0 s). C is solved with k1 + k2 exactly as milestone 4p solved B (sizes 5, 6, 7 and 13, cold start 1.3× C's own k1 Zhang f). C's twin is milestone 4r's (pool_twin, imported unchanged; resid stream index unit + 4, jitter stream tag 7 as in 4r). The scored cells are the ten that hold C: CC, AC, CA, BC, CB on each unit, 7 + 6 design, milestone 4g's 1,716 split pairs, milestone 4i's per-split 99th-percentile threshold, ungated. The A and B arms must reproduce milestone 4r's figures (jittered fr3 BB S0 0.58%, resid-only pooled ±2% 9.08%) before anything is scored.
- H108 (the jitter keeps fresh views safe). Twin with each pool's own fitted τ (4r's rule, C's τ from C's own 6-view solves): S0 false alarms at most 2.5% in each of the ten C cells. Refuted above in any.
- H109 (near nominal). The same, pooled over the ten cells: at most 1%. Refuted above.
- H110 (a borrowed τ is enough). C's twin jittered with the mean of A's and B's fitted τ (as % of f) instead of its own: S0 at most 2.5% in each of the ten cells. Refuted above in any.
- H111 (the jitter is needed). Resid-only twin (τ = 0 everywhere): S0 above 2.5% in at least one of the ten cells. Refuted if every cell is at or below 2.5%.
- H112 (calibrations agree within the jittered twin's sd). 13-view f of C against A and against B, in the jittered twins' sd (4r's std over draws, combined in quadrature): at most 2 sd for all four gaps (two units × two pools). Refuted above 2 on any.
- Reported, unscored. C's fitted τ and the real and twin 6-view spread; per-cell S0 and ±2%, own τ against resid-only (power ratio) and borrowed τ; ±1% and ±4%; milestone 4q's gate (1% cut) on the jittered set sd; 4r's f + principal-point variant; C's median tilt against A and B; anchors against TUM's published f.
- Caveat set in advance. C comes from the same hand-held sequence as A and B, interleaved in time, so this is fresh views of one session, not a new session or mount. Splits share views within a pool; no interval. A pool with fewer usable frames per block after two exclusions may be weaker than A and B for that reason alone; its tilt is reported, not adjusted for.
2026-11-16 · Milestone 3b result: from natural-feature tracks, one 1.2-second window of video gives f to within 3–7% (median), and the solver's sd is 4–12× too small (rms z 9.3, H5 refuted); the median over many windows lands within 1.5% of the board calibration
- Runs.
experiments/m3b_natural/run.py on TUM RGB-D fr1/desk, fr1/xyz (Kinect) and fr3/structure_texture_far (Xtion), fetched by fetch.sh (the TUM host answered this session): 65 disjoint 12-keyframe windows, each solved from 3 starting f (195 solves, about 45 min on 4 cores; results.json); analyse.py scores them (summary.json, natural.png). tests/test_m3b.py checked first that the projection matches cv2.projectPoints, that the undistortion inverts it, that union-find tracks drop components holding two features of one view, and that on a synthetic 8-view scene with 0.5 px noise the solver recovers σ (0.5 ± 0.1 px) and f within 4 of its own sd from an 8% start. - Counted. 57 of 65 windows (fr1/desk 15 of 17, fr1/xyz 18 of 22, fr3 24 of 26). Failures: 6 not converged in 400 iterations, 1 could not initialise (fr1/desk, motion blur), 1 with an undefined covariance (negative variance from a near-singular reduced system; this cause was not in the plan, and the rule
sd_undefined was added to analyse.py after the run). Median 1,034, 1,800 and 2,122 tracks per window, inlier rms 0.55, 0.44 and 0.35 px; 562,769 observations over the counted windows. - H104 refuted. On fr1/desk the median |f − f<sub>ref</sub>| / f<sub>ref</sub> is 2.80% (bar 2%). fr1/xyz 5.18%, fr3 7.24%. Single windows range from −13.7% to +36.5%.
- H105 (= H5) refuted. rms z over the 57 counted windows 9.31 (bar 2): fr1/desk 6.85, fr1/xyz 9.81, fr3 10.22 (against TUM's calibration 6.82, 9.69, 10.59). 2-sd coverage 27%, 11%, 17%.
- H106 refuted. Robust spread of f across windows over the median sd<sub>BA</sub>: 4.20 (fr1/desk), 8.38 (fr1/xyz), 11.74 (fr3); bar 2. The scatter is 3.3%, 9.2% and 11.2% of f; the solver's median sd 0.85%, 1.09% and 0.91%. So the overconfidence is not the reference's fault.
- H107 refuted. Median sd<sub>BA</sub>/f on fr1/xyz is 1.28× that on fr1/desk (bar 2×), while the real scatter is 2.8× larger. The median rotation per window is 23° on fr1/desk, 7.9° on fr1/xyz and 4.8° on fr3 (motion capture); the solver's sd does not follow it (Spearman sd against rotation 0.04, 0.18, 0.18, none significant), while on fr3 the error does (|rel| against rotation +0.58, p = 0.003: the windows that rotate more are further off, not closer).
- Unscored, chosen after scoring. Local minima are common (the three starts end more than 1% apart in 5, 10 and 16 windows) but do not explain the overconfidence: windows whose starts agree have rms z 7.37, 11.36 and 7.19. The median of a sequence's windows is close: fr1/desk 522.0 px (+0.72% against Zhang's 518.28, +0.99% against TUM), fr1/xyz 511.0 (−1.41%, −1.14%), fr3 538.4 (+1.45%, +0.20%); with an sd taken from the across-window spread (1.2533 × MAD / √n) their z's are 0.67, −0.52 and 0.50 (TUM: 0.91, −0.42, 0.07). Median k<sub>1</sub>, k<sub>2</sub>: fr1/desk 0.136, −0.224 (Zhang 0.223, −0.568); fr3 0.002, 0.008 (TUM undistorts fr3, so near 0 is right).
- Reading. H5, the proposal's question for natural features, fails clearly: on hand-held video a 12-keyframe window's covariance understates its focal error 4–12×, more than the 2× milestone 3 found with board corners, and the solver does not notice when the motion is weak (fr1/xyz is nearly pure translation, the critical motion for f, yet its sd is barely larger). What survives is the across-window spread: the median of 15–24 windows lands within 1.5% of both references and the spread gives it an honest-looking interval. For a robot this says not to trust a single short window's sd at all, and to calibrate from the scatter of many windows instead, which is milestone 4r's lesson (a per-view error the model does not carry) in a stronger form. Candidate causes, none tested yet: SIFT localisation errors correlated along a track over consecutive 30 Hz frames, rolling shutter on both sensors during hand-held motion, and the near-critical motion of a 1.2 s window.
- Caveat. As set in advance: windows of one sequence share a scene and a session, and consecutive windows are about a second apart, so the across-window figures are not independent draws; the references differ by 0.3% (fr1) and 1.2% (fr3). The median-of-windows result is post hoc and descriptive.
2026-11-16 · Milestone 3b planned: with natural-feature tracks instead of board corners, is target-free self-calibration still accurate, and is its covariance still honest (H5)?
Written before any scored milestone-3b quantity was computed. Every milestone so far fed the solver a checkerboard's corners (board geometry withheld). The proposal's H5 asks the question a field robot faces: with SIFT tracks of an ordinary scene (mismatches, looser localisation, points seen in only some views), does the z-score of f against the target calibration keep rms ≤ 2 after a robust loss? The TUM host answered this session's proxy, so the data are three TUM RGB-D sequences (CC BY 4.0) shot by the two cameras whose board views milestones 4j–4r used: fr1/desk (Kinect, hand-held sweep over a desk, rotation and translation), fr1/xyz (Kinect, translation along the three axes, little rotation: near the critical motion for focal length) and fr3/structure_texture_far (Xtion, a sideways pass over a textured zig-zag structure; TUM ships fr3 images already undistorted).
- Method (
experiments/m3b_natural/nfba.py, run.py). Disjoint windows of 12 keyframes, every 3rd frame (36 frames, ~1.2 s), from frame 0: 17, 22 and 26 windows. SIFT (OpenCV, 4,000 features, contrast 0.02); matches between keyframes 1, 2 and 3 apart (Lowe ratio 0.75, then a RANSAC fundamental matrix at 2 px); tracks by union-find, one feature per view, ≥ 3 views. Start: essential matrix of the first keyframe and its widest-parallax partner at the starting f, then PnP. Bundle adjustment over f, c<sub>x</sub>, c<sub>y</sub>, k<sub>1</sub>, k<sub>2</sub> (milestone 4n's model), every pose and point, soft-L1 loss with a 1 px knee, Levenberg–Marquardt with IRLS weights and a Schur complement; gauge: first keyframe fixed, scale by one translation component of the farthest keyframe. Covariance of the intrinsics: the IRLS-weighted Gauss–Newton matrix at the solution, points eliminated, σ<sup>2</sup> = weighted SSR / dof. Each window is solved from three starting f (450, 554.3 = a generic 60° field of view, 680 px) and the lowest robust cost is kept. A window counts if it has ≥ 100 tracks, initialises, converges (relative cost change < 10<sup>−7</sup> within 400 iterations), has inlier rms < 1.5 px and 300 < f < 1000; the rest are reported as failures, by cause. - Reference. Primary: each unit's target calibration from milestone 4n, Zhang with k<sub>1</sub> + k<sub>2</sub> on its 13 board views (fr1 518.28 px, fr3 530.71 px). Secondary, reported: TUM's published calibration (mean of f<sub>x</sub>, f<sub>y</sub>: fr1 516.9, fr3 537.3). z = (f − f<sub>ref</sub>) / sd<sub>BA</sub>.
- H104 (accuracy on general motion). On fr1/desk, the median over counted windows of |f − f<sub>ref</sub>| / f<sub>ref</sub> is at most 2%. Refuted above.
- H105 (= H5, honest covariance after a robust loss). Over the counted windows of all three sequences, rms z ≤ 2. Refuted above.
- H106 (honest without the reference). In each sequence, the robust spread of f across its counted windows (1.4826 MAD) is at most 2× the median sd<sub>BA</sub>. Refuted in any sequence above 2. (Separates an overconfident sd from a bias of the reference or of the corner localiser, which H105 cannot.)
- H107 (the solver sees weak motion). The median sd<sub>BA</sub>/f on fr1/xyz is at least 2× that on fr1/desk. Refuted below 2.
- Reported, unscored. Per sequence: counted share and failure causes; median f, relative error to both references, rms z and 1σ/2σ coverage; how often the three starts end more than 1% apart in f and which had the lowest cost; tracks, observations, inlier rms; each window's rotation and translation from TUM's motion-capture ground truth (never shown to the solver) against its sd<sub>BA</sub>.
- Pilot, disclosed. Before this entry, three single-window pilots (one per sequence, windows at frames 0 and 300) were run to make the pipeline converge: a 6-frame step left fr1/desk too few tracks; SciPy's TRF stalled, so the solver became the LM above; holding a component of keyframe 1's translation for scale converged slowly, so the farthest keyframe holds it. One fr3 pilot window showed that two gauges ended in different minima (different cost), which is why three starts are used, and printed sd/f of 1.6% and 0.3% for those two minima; no pilot f was compared with a reference.
- Caveat set in advance. Windows of one sequence share a scene and a capture session, so they are not independent draws of the camera's error; consecutive windows are about a second apart. The references disagree with each other by 0.3% (fr1) and 1.2% (fr3), which bounds how small a z can be read meaningfully, hence H106.
2026-11-16 · Milestone 4r result: one per-view focal jitter, fitted from each pool's own 6-view solves, brings every cell's false alarms under 0.6% (the flat Xtion pool from 4.44%), but halves the test's power; and milestone 4p's "2.7-sd gap" between the Xtion's two calibrations was mostly an sd computed the wrong way
- Runs.
experiments/m4r_jitter/analyse.py (about 2 min on 4 cores, no new solve) rebuilds milestone 4p's twins draw for draw, adds the per-view focal jitter at the fitted τ, and reapplies milestone 4i's per-split threshold, 4p's decide and 4q's gate unchanged (summary.json, jitter.png). Its τ = 0 arm reproduces milestone 4q's ungated figures exactly (fr3 BB S0 4.44%, pooled ±2% 9.08%), which the script asserts. tests/test_m4r.py checked first that a common focal shift moves every subset's twin f by exactly that shift (g<sub>S</sub> · ∂u/∂f = 1 at 5, 6 and 13 views), that the jitter is one draw per view, and the fit of τ on a known case. - Fitted τ (sd of each view's own f, % of f): fr1 A 0.75%, B 1.88%; fr3 A 0.22%, B 1.80%. The twin's 6-view robust spread without it: 0.45%, 0.80%, 0.46%, 0.63% against real 0.59%, 1.23%, 0.50%, 1.23%, so the flat pools (B) need the most.
- H100 held. Within B on fr3, S0 false alarms 0.58% (from 4.44%).
- H101 held. Every cell at or below 0.58%: fr1 AA 0.29%, BB 0.35%, AB 0.58%, BA 0.29%; fr3 AA 0.12%, BB 0.58%, AB 0.35%, BA 0.41%. Pooled 0.37% (resid-only 1.42%), now below the 1% nominal level: the jittered threshold is conservative.
- H102 refuted. Pooled ±2%: 4.07% against 9.08% (0.45×; bar 0.8×). Per cell fr1 AA 14.0% (19.3%), BB 0.8% (2.1%), AB 2.0% (4.9%), BA 2.1% (5.4%); fr3 AA 7.6% (10.3%), BB 1.9% (9.2%), AB 2.1% (9.2%), BA 2.1% (12.3%). Median per-split threshold on z<sub>R</sub><sup>2</sup> rises from 10–14 to 12–35 (fr3 BB 11.8 to 35.2).
- H103 held, but not because of the jitter. A-against-B gap in 13-view f, in the jittered twin's sd: fr1 0.15, fr3 0.54. With the resid-only twin's sd taken the same way (std over its own draws) the gap is already 0.29 and 1.59. Milestone 4p's 0.98 and 2.71 used ‖g<sub>13</sub>‖ × rms, the iid-noise sd, which on these units is 1.6–3.5× smaller than the resid model's own (fr3 A 0.74 against 1.37 px; fr1 B 0.77 against 2.73 px), because the resid model's within-view correlation does not average away over 13 views. So the "two calibrations of the Xtion disagree by 2.7 sd" reading of milestones 4p and 4q is withdrawn: under the twin's own noise model they agree within 1.6 sd.
- Unscored. With milestone 4q's gate on the jittered set sd (1% cut): 15.5% of splits scored (resid-only gate 44%), S0 0.28% pooled, worst cell 0.54%, ±2% 15.9% of scored splits (4q's gate 15.9% at 44% scored), ±4% 61.3%. ±1% ungated 1.2% (resid-only 3.4%), ±4% 17.8% (31.0%). Jittering cx and cy as well as f at a jointly fitted τ (0.28%, 0.65%, 0.10%, 0.60% of f) is between the two: pooled S0 0.68%, worst cell fr3 BB 1.81%, ±2% 5.9% (0.65×). The predicted-power figure listed in the plan was not computed (see the run record).
- Reading. A single number per pool, read off the scatter of the robot's own small-set solves, is enough to make the update test safe on every pool and pairing here, including the flat Xtion pool that defeated milestones 4p and 4q, so the over-alarm was the twin understating how much each view's solve moves, as milestone 4q concluded. The price is steep: the threshold widens so much that the test catches a 2% focal change half as often (4.1% of splits), and with false alarms at 0.37% against a 1% target it is now too cautious, because a per-view f offset is one of several shapes the missing error could take and the one that costs the most power (the f + principal-point variant keeps 0.65× of the power at 1.8% worst-cell false alarms). Combined with the gate, the robot tests rarely (15% of splits) but then catches 2% changes as often as before (15.9%) at 0.3% false alarms. The 13-view check H103 corrects an earlier reading rather than confirming this one: the Xtion's two calibrations were never as far apart as milestone 4p said.
- Caveat. As set in advance: the pools are 4p's and 4q's, whose results motivated the term, so these passes are consistency checks, not confirmation on unseen data; τ summarises whatever per-view error there is and does not identify it. Splits share views, so no interval.
2026-11-16 · Milestone 4r planned: does a per-view focal jitter, fitted from the pool's own solves, make the update test's threshold safe?
Written before any milestone-4r quantity was computed, after milestone 4q's results were read. Milestone 4q found that the flat Xtion pool's over-alarm (fr3 B, S0 4.44% ungated, 5.85% gated) is not the far solves but splits the twin rates as precise and are not, and read that as an error shared within a view that the twin's noise (each view's 13-view residual pattern, resampled) does not carry. Two figures already on file point the same way: the real robust spread of 6-view f is 1.23% on fr3 B against a median twin sd of 0.65% (milestones 4p, 4q), and the two 13-view calibrations of the Xtion differ by 2.71 of the twin's sd (4p). This milestone adds the simplest such error to the twin: each view sees its own effective focal length, f + δf<sub>v</sub>, δf<sub>v</sub> ~ N(0, τ<sup>2</sup>) independent across views (the physical stand-ins: rolling shutter and motion during a hand-held sweep, frame timing, board flex), entered as the corner shift ∂u/∂f · δf<sub>v</sub> at the twin. τ is fitted per pool from that pool's own solves only (what a robot has): the value at which the twin's robust spread of 6-view f across sets (1.4826 MAD over the 1,716 sets, median over draws) equals the real one (milestone 4n / 4p's spread, n = 6); τ = 0 if the twin already matches. Fitted on the first 2,000 of each pool's 10,000 draws; the resid draws are milestone 4p's (same seeds and streams), the jitter a separate stream (seed 20261114, tag 7). Everything else is milestone 4p / 4q's code imported unchanged: twin sets, milestone 4i's per-split 99th-percentile threshold on z<sub>R</sub><sup>2</sup>, the 1,716 split pairs of the 7 + 6 design, the four pairings AA, BB, AB, BA. The resid-only arm must reproduce milestone 4q's ungated figures (fr3 BB S0 4.44%, pooled ±2% over all four pairings 9.1%) before anything is scored.
- H100 (the fitted jitter makes the flat Xtion pool safe). Within B on fr3, ungated S0 false alarms at most 2.5% (resid-only 4.44%). Refuted above.
- H101 (and every cell). Ungated S0 at most 2.5% in each of the eight unit × pairing cells. Refuted above in any.
- H102 (at a small cost in power). Pooled ±2% (over sign, unit and the four pairings) at least 0.8× the resid-only pooled rate. Refuted below.
- H103 (the fit predicts what it was not fitted to). With the jittered twin's sd of each pool's 13-view f (std over its draws), the A-against-B gap in 13-view f is at most 2 sd on both units (resid-only: 0.98 on fr1, 2.71 on fr3). The 13-view f's were not used in the fit. Refuted above 2 on either.
- Reported, unscored. Fitted τ per pool (as % of f); per-cell S0 and ±2%; ±1% and ±4%; milestone 4q's gate (1% cut) recomputed with the jittered twin's set sd; a variant that also jitters the principal point per view (δc<sub>x</sub>, δc<sub>y</sub> at the same fitted τ); the jittered twin's mean predicted power.
- Caveat set in advance. The pools are milestones 4p and 4q's, whose results motivated the term, so a pass is a consistency check on the data that suggested it, not a confirmation (a third disjoint pool is still the test; the TUM host refused this session's proxy in milestone 4q). τ is a one-number summary of whatever per-view error there is; the experiment does not identify its physical cause. Splits share views, so no interval.
2026-11-16 · Milestone 4q result: abstaining when the twin says the views cannot resolve 2% raises the update test's power 1.75× and removes every far solve, but does not make the flat Xtion pool safe (5.9% false alarms): the over-alarm is not the far solves
- Runs.
experiments/m4q_gate/analyse.py (about 6 min on one core, no new solve) rebuilds milestone 4p's twins and per-split thresholds with its code imported unchanged, adds each set's twin sd of f, and scores the gate (summary.json, gate.png). It reproduces milestone 4p first (within-B fr3 S0 4.44%, ±2% across pools 7.96%), which the script asserts. tests/test_m4q.py checked first the set sd, the far flags and that gated rates count only scored splits. - H96 refuted. Within B on fr3, gated S0 false alarms 5.85% (564 of 1,711 splits scored), up from 4.44%.
- H97 refuted. Gated S0 per cell: fr1 AA 0.79%, AB 0.22%, BA 0.17%, BB none scored; fr3 AA 0.08%, AB 2.27%, BA 1.93%, BB 5.85%.
- H98 held. Gated ±2%, pooled over sign, unit and the four pairings: 15.9% against 9.1% ungated (×1.75). Per cell the gated rate is above the ungated one in every scored cell (fr1 AA 25.2% against 19.3%, AB 7.9% against 4.9%, BA 9.2% against 5.4%; fr3 AA 13.1% against 10.3%, BB 12.3% against 9.2%, AB 14.7% against 9.2%, BA 18.3% against 12.3%).
- H99 refuted. Scored share: fr1 AA 74%, AB 27%, BA 35%, BB 0%; fr3 AA 74%, BB 33%, AB 52%, BA 57%.
- Unscored. Median twin sd of a 6-view set's f: fr1 A 0.49%, B 1.09%; fr3 A 0.48%, B 0.65% of f, so the gate reads milestone 4p's "B is weaker" from the twin alone. Splits holding a far solve (> 5% from its size's median): 4, 42, 34, 12 (fr1 AA, BB, AB, BA) and 20, 149, 112, 57 (fr3); the gate scores none of them. Cut-off sweep, pooled (scored share, S0, ±2%, worst cell's S0): 0.75% 22%, 1.24%, 23.2%, 3.1%; 1% 44%, 1.41%, 15.9%, 5.9%; 1.5% 74%, 1.51%, 11.7%, 6.1%; 2% 89%, 1.46%, 10.1%, 5.3%. ±1% at the 1% cut 5.0% (ungated 3.4%), ±4% 54.4% (31.0%). A tilt-only gate (new set at least as tilted as the old): 53% scored, S0 1.50%, ±2% 10.1%, worst cell fr3 BB 5.3%.
- Reading. One prediction held and it is the one a robot would use first: a gate that needs nothing but the twin it already has removes every split with a far solve and buys three quarters more power on the splits it keeps, and on the Kinect it turns the flat pool's weak test (2.1% power) into an abstention rather than a guess. What it does not do is make the Xtion's flat pool safe, and that overturns milestone 4p's reading: the 4.4% over-alarm is not the far solves the linear twin cannot represent (none survive the gate, and the rate rises), but splits the twin rates as precise and are not. That fits two other numbers: the two 13-view calibrations of the Xtion differ by 2.7 of the twin's own sd (milestone 4p), and on the Xtion the twin overstates the test's power 1.96× (milestone 4o). Both say the twin's per-set sd is too small on this unit by an error shared within a view, which neither the corner noise nor the residual pattern carries. The safe reading for a robot today: gate on the twin sd to choose when to test, and size the threshold for a per-view error the twin does not yet model.
- Caveat. As set in advance: the pools are milestone 4p's and its over-alarm motivated the gate, so this checks the gate on the data that suggested it; a third disjoint pool (the TUM host refused this session's proxy) is still the test. Splits share views, so no interval; the sweep is descriptive.
2026-11-16 · Milestone 4q planned: can the robot tell from the views alone when not to run the update test?
Written before any milestone-4q quantity was computed, and after milestone 4p's results were read (so this is not a test on unseen data; see the caveat). Milestone 4p found that a flatter pool of views (B) makes the 7 + 6 update test both weaker and, on the Xtion, unsafe within the pool (S0 false alarms 4.44%), and its next step asked for a reference or gate built from the new set's own geometry, checked on a third disjoint pool. The TUM host refused this session's proxy (403, probed 2026-11-16), so no third pool can be extracted; this milestone scores a gate on milestones 4n and 4p's stored solves of pools A and B instead. The gate uses only what a robot has before it compares anything: the k1 + k2 twin of each pool (milestone 4p's twin_pool, imported unchanged, same draws and seeds) gives every view set's first-order sd of f, sd<sub>S</sub> = (sd of the twin's f of S over its 10,000 resid draws) × the pool's 13-view fit rms. A split (old set o from one pool, new set n from the same or the other) is scored only if √(sd<sub>o</sub><sup>2</sup> + sd<sub>n</sub><sup>2</sup>) ≤ 1% of the old pool's 13-view f, i.e. if the design target (a 2% focal change) sits at least two of the twin's own sd's from zero; otherwise the test abstains. The 1% is fixed here from that target, not fitted; other cut-offs (0.5% to 2%) are reported unscored. Rates below are over scored splits, on milestone 4g's 1,716 split index pairs, with milestone 4p's per-split twin threshold, in all four pool pairings (AA, BB, AB, BA). The ungated figures must reproduce milestone 4p's (within-B fr3 S0 4.44%, pooled ±2% across pools 7.96%) before anything is scored.
- H96 (the gate makes the flat Xtion pool safe). Within B on fr3, gated S0 false alarms at most 2.5% (ungated 4.44%). Refuted above.
- H97 (and keeps every pairing safe). Gated S0 false alarms at most 2.5% in each of the eight unit × pairing cells. Refuted above in any.
- H98 (abstaining buys power). Gated ±2% rate, pooled over sign, unit and the four pairings, at least 1.2× the ungated pooled rate. Refuted below.
- H99 (the gate still lets the robot test). At least 20% of usable splits scored in each of the eight cells. Refuted below in any.
- Reported, unscored. Per-cell scored shares, S0 and ±2% rates; ±1% and ±4%; the cut-off sweep; how many splits holding a far solve (more than 5% from the median of its size) the gate drops; the median sd<sub>S</sub> of 6-view sets per pool; a tilt-only gate (scored only if the new set's median tilt measure is at least the old set's) for comparison.
- Caveat set in advance. The data are milestone 4p's and its within-B over-alarm motivated the gate, so a pass here is a consistency check on the pools that suggested it, not a confirmation; a third disjoint pool (or another sequence) is still the test. Splits share views, so no interval.
2026-11-16 · Milestone 4p result: on 13 views it has never seen, the update test stays under the 2.5% false-alarm bar (0.9% and 1.9%) but not at nominal (1.4% pooled), and catches a 2% focal change in 8.0% of splits against 14.8% in-sample; the fresh views are less tilted, and the drop follows the weaker pool, not freshness alone
- Runs.
experiments/m4p_fresh/prepare.py (detection on all 1,507 + 1,327 frames, a few minutes) reproduced milestone 4j's 13 frames per unit and picked a disjoint 13; run.py solved 9,440 k1 + k2 subsets of the fresh pools (about 40 min on 4 cores); analyse.py (about 5 min) builds a twin per pool and scores the 7 + 6 test within and across pools (summary.json, fresh.png). Pool A reproduces milestone 4n's k1 + k2 figures exactly (S0 0.588% / 0.117%, ±2% pooled 14.76%), which the script asserts before scoring. tests/test_m4p.py checked first that the pools are disjoint and at least 1 s apart, the pick rule's exclusion, and that the cross-pool decide equals milestone 4g's decisions within one pool. - H92 held. S0 across pools, both directions: 0.91% (fr1; A old 1.17%, B old 0.65%) and 1.90% (fr3; 1.87% and 1.93%), against 0.59% and 0.12% within A.
- H93 refuted. Pooled: 1.40% (bar 1%).
- H94 refuted. ±2% across pools, pooled over sign, unit and direction: 7.96%, 0.54× the in-sample 14.76% (bar 0.7×). fr1 4.9% / 5.4% (A old / B old) against 19.3% within A; fr3 9.2% / 12.3% against 10.3%.
- H95 refuted, narrowly, on fr3. 13-view k1 + k2 f of B against A: fr1 526.04 against 525.16 (+0.17%), fr3 537.14 against 534.04 (+0.58%; bar 0.5%). In the pools' own twin sd of the 13-view f (resid model at each fit's rms) the gaps are 0.98 (fr1) and 2.71 (fr3).
- Unscored: pool B is a weaker set of views. Median tilt of the kept frames (prepare's foreshortening measure) 0.147 against 0.268 (fr1) and 0.063 against 0.101 (fr3); spacing the same (37 against 37 px, 34 against 35). Robust spread of f over 6-view sets 1.23% against 0.59% (fr1) and 1.23% against 0.50% (fr3); far 6-view solves (> 5% from the median) 34 against 6 and 112 against 19. Within B: S0 false alarms 0.58% (fr1) and 4.44% (fr3), above the programme's 2.5% bar; ±2% 2.1% (fr1) and 9.2% (fr3). Pooled ±2%: A 14.8%, B 5.7%, across 8.0%, so the across-pool rate sits between the two pools' own rates rather than below both. Median per-split threshold: fr1 10.3 (A), 14.2 (B), 13.2 / 13.1 across; fr3 10.7, 11.7, 11.4 / 12.0. The plain jackknife J across pools at S0: fr1 0.82% / 0.65%, fr3 1.75% / 1.28%; χ²<sub>1</sub>: 2.7% / 2.7%, 5.0% / 5.3%. ±4% across pools, pooled over sign and direction: fr1 23.5%, fr3 33.5% (within A 53.6% and 46.9%). Anchors against TUM's published f: target-free fr1b +1.77% (fr1 +1.60%), fr3b −0.03% (fr3 −0.61%); flat-board Zhang k1 + k2 on fr1b 523.42 (+1.2%) and on fr3b 545.01 (+1.43%, against fr3's 530.71, −1.2%), so on fr3b's flatter views the target-free solve lands nearer TUM than Zhang does.
- Reading. Three of four predictions failed, and the way they failed is the result. On views the test never saw, the twin reference still keeps false alarms below the 2.5% bar (1.4% pooled), but the in-sample 0.1–0.6% was optimistic by a factor of 1.5 (Kinect) to 16 (Xtion); the power halves. Most of that is not freshness but geometry: the rule that picked the fresh frames was the same, yet it found flatter boards, and on a flatter pool the solves scatter twice as much and the test, within that pool alone, both loses power (2.1% on the Kinect) and, on the Xtion, over-alarms (4.4%) because far solves the linear twin cannot represent sit in its tail. The two 13-view calibrations of the Xtion differ by 2.7 of the twin's own sd, so the twin also understates the between-pool variation of the full calibration there. For a robot this says the test's safety depends on the views as much as on the threshold: a new set that is flatter than the one the threshold was built on should be held back or scored with a reference built on its own views, and an in-sample false-alarm rate should be read as a lower bound.
- Caveat. As set in advance: one hand-held sequence per unit, B interleaved in time with A, so this is fresh views of one session, not a new session or mount; splits share views within a pool, the two directions are not independent, no interval. The tilt difference between the pools was seen before any solve (the cold-start Zhang f was printed) and the hypotheses were left as written; it is a description of the pools, not a fitted explanation, and no experiment here separates the effect of tilt from that of freshness.
2026-11-16 · Milestone 4p planned: does the update test keep its false-alarm rate and its power on views it has never seen?
Written before any milestone-4p solve or statistic was computed. Every update-test figure since milestone 4g scores splits of the same 13 views per unit, so the old and new sets, the twin and its reference all come from one pool (paper/draft.md, Section 9: scoring every split "says nothing about a fresh set of views"). Here a second, disjoint set of 13 views is picked from each TUM sequence (experiments/m4p_fresh/prepare.py): the same detection, spacing floor, 13 time blocks and pick rule as milestone 4j's prepare.py, imported unchanged, except that in each block the frame 4j kept and every frame within 1.0 s of it are removed first. The fresh set B is solved with k1 + k2 (milestone 4n's solver, start 1.3× B's own k1 Zhang f, sizes 5, 6, 7 and 13: 4,720 solves per unit), its own k1 + k2 twin is fitted, and the 7 + 6 test is scored with the old set from one pool and the new set from the other, in both directions (A old, B new; B old, A new), on milestone 4g's 1,716 split index pairs, against the per-split twin threshold built from the two pools' twins (milestone 4i's references, as in the swap scenario S3; independent draws per pool, seed 20261114, A's draws the ones milestone 4n used). The in-sample comparison is milestone 4n's k1 + k2 7 + 6 figures, which A must reproduce first (S0 0.59% / 0.12%, ±2% pooled 14.8%).
- H92 (fresh views keep the false alarms safe). S0 across pools, k1 + k2, 7 + 6, both directions pooled: R above its per-split twin threshold on at most 2.5% of splits on each unit. Refuted above on either.
- H93 (and near nominal). The same, pooled over units and directions: at most 1%. Refuted above.
- H94 (the power carries over). ±2% applied to the new set, pooled over sign, unit and direction: at least 0.7× the in-sample k1 + k2 rate (milestone 4n: 14.8%), so at least 10.4%. Refuted below.
- H95 (two disjoint calibrations of one camera agree). 13-view k1 + k2 f of B against A: within 0.5% on each unit. Refuted on either.
- Reported, unscored. Per direction and unit rates; ±1% and ±4%; the in-sample S0 and ±2% rates within B (B's own 1,716 splits, the same code as milestone 4n); the plain jackknife J and χ²<sub>1</sub>; B's anchors against TUM's published f and flat-board Zhang on B; the median per-split threshold across pools against within a pool; converged counts and far solves.
- Seen after the plan, before any target-free solve.
prepare.py reproduced 4j's 13 frames on both units and picked B's 13 at 1.0–24.3 s (fr1b) and 1.2–15.6 s (fr3b) from 4j's frames. The cold starts print the k1 flat-board Zhang f of each pool: fr1 520.79, fr1b 524.47 (+0.71%); fr3 530.59, fr3b 544.85 (+2.69%). fr3b's views are less tilted than fr3's (tilt measure 0.04–0.23, nine of 13 below 0.07), so H95 may fail on fr3 for a reason the plan did not name; the hypotheses are left as written. - Caveat set in advance. B comes from the same hand-held sequence as A, interleaved in time (one view per block each), so it is fresh views of the same session, not a new session, temperature or mount; a drift across sessions is not tested. Within a pool the splits still share views, so no interval is given; the two directions are not independent of each other.
2026-11-15 · Milestone 4o result: with k1 + k2 on the Kinect the twin's predicted power matches the real rate (0.995×, from 1.34×), and the gains compound: fitting k2 and then waiting for the twin's best new set catches a 2% focal change in 17.9% of old sets against 6.2% blind with k1 (×2.86)
- Runs.
experiments/m4o_k2wait/analyse.py (about 3 min on 4 cores) builds a k1 twin on milestone 4j's solves and a k1 + k2 twin on milestone 4n's, from the same 10,000 resid draws per unit, and runs milestone 4l's predict on 7 + 6 and 6 + 6; score.py (under a minute) applies milestone 4l's unit_stats and 4m's design_stats and pooled, imported unchanged (summary.json, k2wait.png). The k1 arm reproduces milestones 4l and 4m's TUM figures exactly (fr1 AUC 0.749, mean π / real 1.34, median threshold 9.52; 6 + 6 pooled chosen 9.73% against 7.33% blind), checked before the k1 + k2 arm was read. tests/test_m4o.py checked first σ̂'s parameter count with the extra lens term (p 219 to 220), that the twin sets restricted to sizes 5–7 give the same R as the full ones for 6- and 7-view sets, and the pooled top-quarter rate and the H90 ratio on known cases. - H86 held. AUC of π<sub>s</sub> for the real ±2% detection at 7 + 6, k1 + k2: fr1 0.775 (k1 0.749), fr3 0.749 (0.748).
- H87 held. Mean π over the real ±2% rate on fr1: 0.995 (0.192 against 0.193), against k1's 1.34. fr3 (unscored): 1.96 against 1.89.
- H88 held. Top quarter by π, TUM pooled: 30.7% against 14.8% over all splits (×2.08; k1 21.5% against 10.0%). fr1's top quarter 38.9% (k1 20.3%), fr3's 22.6% (unchanged).
- H89 held. 6 + 6, TUM pooled, old set fixed: the twin's pick catches ±2% in 15.2% of old sets against 11.3% blind (×1.35; k1 ×1.33).
- H90 held. The k1 + k2 chosen rate is 1.56× the k1 chosen rate (15.2% against 9.73%).
- H91 held. S0 false alarms, k1 + k2: top quarter at 7 + 6 1.65% (fr1; k1 2.35%) and 0.23% (fr3); chosen splits at 6 + 6 0.59% and 0.35%.
- Unscored. 6 + 6 under k1 + k2, which milestone 4n did not score: blind ±2% on fr1 14.0% (k1 6.2%), fr3 8.6% (8.4%); S0 false alarms of a blind pick 0.37% and 0.34% (k1 0.63% and 0.13%). Per unit, k1 blind to k1 + k2 chosen: fr1 6.2% to 17.9% (×2.86), fr3 8.4% to 12.5% (×1.49); pooled 7.33% to 15.2% (×2.07). Ceiling (any of the 7 candidates fires) on fr1 32.7% (k1 19.2%), so the twin's pick reaches 55% of it (k1 42%); fr3 26.1% (unchanged), 48%. Within-old-set AUC at ±2%: 0.710 and 0.702 (k1 0.640 and 0.681). Predicted against real gain at 6 + 6: fr1 1.31 against 1.28, fr3 1.25 against 1.46. The sd rule: 15.6% pooled (π's 15.2%), the same pick as π in 42% and 41% of old sets. ±4% at 6 + 6: fr1 chosen 52.3% (k1 31.6%); ±1%: 5.1% (1.8%). At 7 + 6 every split holding a far solve sits in the bottom quarter by π (fr1 4 of 4, fr3 20 of 20; k1 8 and 36), and the sd ranks as well as π (AUC 0.775 and 0.746).
- Reading. The two gains are independent and multiply. Fitting the lens order the Kinect needs doubles the update test's power at 6 + 6 as it did at 7 + 6 (6.2% to 14.0% blind), and the twin's choice among new sets adds the same third on top (×1.28 on fr1, ×1.35 pooled), so a robot that does both catches a 2% focal change nearly three times as often as one that does neither, with false alarms at or below 0.6%. The twin's level, 1.34× too high on the Kinect with k1, is right with k1 + k2: part of the "real f varies more than the twin says" gap of milestone 4l was distortion the k1 model left to be absorbed differently by each view set, which the linear twin cannot represent. It is not all of it: on the Xtion, whose lens k1 already follows, the twin still overstates its power 1.96×, so the remaining overstatement is something else (per-view effects the residual-pattern noise does not carry). The twin ranks a little better on the Kinect too (AUC 0.749 to 0.775 across splits, 0.640 to 0.710 within an old set).
- Caveat. As set in advance: 13 views per unit, splits and candidates sharing views, no interval; 5 + 6 (28 candidates) not run, its 4-view delete-one solves were never computed with k1 + k2. The k1 + k2 solves are milestone 4n's, untuned.
2026-11-15 · Milestone 4o planned: with the lens the Kinect needs (k1 + k2), does the twin still rank updates, and does waiting for the best new set add to the doubling?
Written before any milestone-4o quantity was computed. Milestone 4n found that fitting k1 + k2 on the Kinect (fr1) doubles the 7 + 6 update test's power for a ±2% focal change, and milestone 4m that, with the old calibration held fixed, picking the candidate new set the twin rates highest gains ×1.33 on the TUM units (k1). The paper's open question (paper/draft.md, conclusion) is whether those two gains compound. No new solve: milestone 4n's k1 + k2 solves (5, 6, 7 and 13 views per unit) cover 7 + 6 and 6 + 6, so milestone 4l's predicted power (rig_modules, sigma_hat, predict, imported unchanged; σ̂ counts the extra lens parameter) and milestone 4m's choice rule (groups, choose, design_stats, imported unchanged) are run with a k1 + k2 twin beside a k1 arm on 4j's stored solves (same 10,000 resid draws per unit, seed 20261114). The k1 arm must reproduce milestones 4l and 4m's TUM figures (fr1 7 + 6 AUC 0.749, mean π / real 1.34; 6 + 6 chosen 9.73% against 7.33% blind) before anything is scored.
- H86 (the twin still ranks on the Kinect). k1 + k2, 7 + 6: AUC of π<sub>s</sub> for the real ±2% detection at least 0.70 on each TUM unit (k1: 0.749, 0.748). Refuted below on either.
- H87 (the twin's level is closer on the Kinect). k1 + k2, fr1, 7 + 6: mean π over the real ±2% rate nearer 1 than k1's 1.34 (|log ratio| smaller). Refuted if not.
- H88 (testing the top quarter still pays). k1 + k2, 7 + 6, TUM pooled: the real ±2% rate in the top quarter of each unit's splits by π is at least 1.5× the rate over all splits. Refuted below.
- H89 (waiting still pays). k1 + k2, 6 + 6, TUM pooled: the chosen rate is at least 1.2× the blind rate (k1: ×1.33). Refuted below.
- H90 (and the gains compound). k1 + k2, 6 + 6, TUM pooled: the chosen rate is at least 1.3× the k1 chosen rate (9.73%). Refuted below.
- H91 (without false alarms). k1 + k2: S0 false alarms at most 2.5% on each unit for the top quarter at 7 + 6 and for the chosen splits at 6 + 6. Refuted above on either.
- Reported, unscored. The 6 + 6 blind rates and S0 false alarms under k1 + k2 (milestone 4n scored 7 + 6 only); ceilings; within-old-set AUC; the sd rule; ±1% and ±4%; the twin's predicted gain against the real one.
- Caveat set in advance. The same 13 views per unit; splits and candidates share views, so no interval. 5 + 6 (28 candidates) cannot be run: its 5-view sets need 4-view delete-one solves that milestone 4n did not compute.
2026-11-15 · Milestone 4n result: fitting the Kinect's lens with k1 + k2 doubles the update test's power for a 2% focal change (9.8% to 19.3% of 7 + 6 splits) at unchanged false alarms; on the undistorted Xtion it changes nothing
- Runs. 9,440 k1 + k2 solves (
experiments/m4n_k2/run.py, 5-, 6-, 7- and 13-view sets of both units, about 2 h on 4 cores, most of it fr1); analyse.py (about 4 min) builds a k1 and a k1 + k2 twin per unit from the same 10,000 resid draws and scores every 7 + 6 split under both. The k1 arm reproduces milestone 4j's 7 + 6 figures exactly (S0 0.763% / 0.058%, ±2% 10.03%, median per-split threshold 9.52 / 10.84, 8 far 6-view solves), which was checked before the k1 + k2 solves finished. tests/test_m4n.py checked first that the k1 + k2 twin reproduces the k1 + k2 solver's 13-view solve and a small nonlinear perturbation, that each private module keeps its own lens model, and that the restricted twin_sets matches milestone 4i's. - H82 refuted. Robust spread of f over 6-view sets, k1 + k2 over k1: 0.81 (fr1, 0.73% to 0.59% of f) and 0.97 (fr3, 0.51% to 0.50%). The extra term did not cost precision; on the Kinect it bought some.
- H83 refuted. fr1 6-view solves more than 5% from the median: 6 (k1: 8). fr3: 19 (k1: 30).
- H84 held. R against its per-split k1 + k2 twin threshold, S0 at 7 + 6: 0.59% (fr1) and 0.12% (fr3); k1 0.76% and 0.06%.
- H85 refuted. ±2% at 7 + 6, pooled: 14.8% against k1's 10.0%. All of the gain is on the Kinect: fr1 21.9% / 16.6% (−2% / +2%) against 11.0% / 8.6%; fr3 9.6% / 11.0% against 10.1% / 10.4%.
- Unscored. Paired, on the 1,696 fr1 splits usable under both models: at −2% the k1 + k2 test alone fires on 241 and the k1 test alone on 53 (agreement 82.7%); at +2% 191 and 53; under S0 8 and 10 (agreement 98.9%). fr3 agrees on 95–96% at ±2% and 99.8% under S0. ±4%: 41.5% to 50.2%; ±1%: 2.0% to 4.1%. S3 (the other unit's old views): 6.3% / 7.9% to 16.3% / 17.0%, with the units' 13-view f now 1.69% apart (k1: 1.40%; TUM: 3.5%). The plain jackknife at ±2%: fr1 13.6% / 9.7% to 21.3% / 18.5%. Median per-split threshold 10.31 (fr1) and 10.68 (fr3). Anchors: target-free 13-view f on fr1 525.16 px with k1 + k2 (+1.60% against TUM's 516.9; rms 0.152 to 0.129 px), fr3 534.04 (−0.61% against 537.3); flat-board Zhang on the same 13 views with k1, k1 + k2 and k1 + k2 + k3: fr1 520.79, 518.28, 517.39 (+0.75%, +0.27%, +0.09%), fr3 530.59, 530.71, 530.67. The one far 5-view k1 + k2 solve on fr1 sits about 175× the median (a 5-view set's own f, used only as a delete-one value).
- Reading. The pre-registered picture was wrong in the useful direction. On a lens that k1 cannot follow, the unmodelled distortion does not sit quietly in the residuals: each view set absorbs a different share of it into f, so the k1 solves of 6-view sets scatter more than the views' noise says, and a 2% focal change hides in that scatter. Adding k2 takes most of it out (spread −19%, far solves 8 to 6) and the update test catches twice as many ±2% changes on the Kinect and more than twice as many real unit swaps (6.3% / 7.9% to 16.3% / 17.0%, the swap involving the Kinect either way), with false alarms unchanged (0.59%). On the Xtion, whose images arrive undistorted, k2 changes nothing measurable: no precision cost, no gain. A fleet should fit the lens order its lens needs, and the cost of one term too many is not visible at 6 views here. The remaining +1.6% bias of the target-free f on the Kinect is not the lens order: flat-board Zhang on the same corners moves from +0.75% to +0.09% of TUM's figure as k2 and k3 are added, while the target-free solve stays at +1.6%, so the gap is the freedom the free board points leave (the board's metric shape traded against f and distortion), which milestone 5 showed is the only difference between the two solvers.
- Caveat. Only the 7 + 6 design; one hand-held sequence of 13 views per unit, splits sharing views, no interval. The k1 + k2 solves and threshold of fr1 were not tuned: the plan's solver, start and seed were used as written. k3 was not tried in the target-free solver.
2026-11-15 · Milestone 4n planned: does a second radial term (k1 + k2) change the update test on the TUM Kinect, whose lens k1 alone cannot follow?
Written before any milestone-4n subset statistic was computed. Milestone 4j's caveat: fr1's lens (a Kinect; TUM's d0 0.26, d1 −0.95, d4 1.16) is not described by k1 alone, its 13-view f sits +1.85% above TUM's, and its residuals carry the unmodelled distortion into the twin. Every milestone since (4k–4m) kept k1 only. Here the same stored corners, solver and cold start are re-run with k1 + k2 (experiments/m4n_k2/run.py, nk = 2, nothing else changed) on the sizes the 7 + 6 design needs (5, 6, 7 and 13 views: 4,720 solves per unit), and milestone 4j's 7 + 6 analysis (splits, scenarios, J and R, the per-split twin reference with the resid model, 10,000 draws, seed 20261114) is run on them with a k1 + k2 twin. fr3 (an Xtion whose images TUM delivers undistorted) is the control: k2 should buy it nothing but cost it precision.
- Seen before this plan. A timing probe printed the 13-view solves with k1 and k1 + k2 (fr1 526.47 to 525.16 px, rms 0.152 to 0.129; fr3 533.87 to 534.04), flat-board Zhang with k1 and k1 + k2 on the same 13 views (fr1 520.79 and 518.28; fr3 530.59 and 530.71), and six k1 + k2 solves at each of 5, 6 and 7 views per unit. These are reported, not scored.
- H82 (k2 costs precision). Robust spread (1.4826 MAD over the median) of f over all converged 6-view sets, k1 + k2 over k1 (milestone 4j: 0.73% fr1, 0.51% fr3): at least 1.2 on each unit. Refuted below on either.
- H83 (and more far solves on the Kinect). fr1 6-view solves more than 5% from the size's median: more than milestone 4j's 8 of 1,709. Refuted if 8 or fewer. (fr3's reported against its 30.)
- H84 (the twin reference stays safe under the richer model). R against its per-split k1 + k2 twin threshold, S0 at 7 + 6: false alarms at most 2.5% on each unit (k1: 0.76% and 0.06%). Refuted above on either.
- H85 (k2 lowers the power). ±2% at 7 + 6, pooled over sign and unit: below milestone 4j's 10.0% (k1, same analysis). Refuted at or above.
- Reported, unscored. The 13-view anchors against TUM, including Zhang with k1 + k2 + k3 (TUM's own model order); the paired comparison on splits usable under both models (how often the ±2% verdict agrees); ±1% and ±4%; S3 (each unit's new views against the other unit's old ones; TUM puts the real f difference at 3.5%) with k1 + k2; the plain jackknife; χ²<sub>1</sub>; converged counts.
- Caveat set in advance. Thirteen views per unit from one hand-held sequence each; the splits share views, so no interval is given. Only the 7 + 6 design (the paper's recommended one) is re-run. k1 + k2 is still not TUM's five-term model, so a residual lens error remains on fr1.
2026-11-15 · Milestone 4m result: with the old calibration fixed, the new set the twin rates highest catches a 2% focal change 1.33× as often as a blind pick on the TUM cameras (7.3% to 9.7%), 1.63× with 28 candidates; the twin gets the gain right but leaves most of the ceiling unused
- Runs.
experiments/m4m_wait/analyse.py --rig opencv and --rig tum (about 15 min each, run side by side on 4 cores): milestone 4l's rig_modules, sigma_hat and predict imported unchanged, on 6 + 6, 5 + 6, 7 + 5 and 7 + 4 (12,012, 36,036, 10,296 and 25,740 splits per OpenCV camera; the TUM units have slightly fewer usable). score.py groups the splits by old set and scores H77–H81 (summary.json, wait.png). tests/test_m4m.py checked first the grouping (old sets with one candidate dropped), the tie rule, the blind rate as the group mean, the within-old-set AUC and the pooling on known cases. Old sets counted: 1,716 (6 + 6) and 1,287 (5 + 6) per OpenCV camera; 1,708 / 1,715 and 1,267 / 1,280 on fr1 / fr3. - H77 refuted. TUM, 6 + 6, pooled: the twin's pick catches ±2% in 9.73% of old sets against 7.33% blind, ×1.33 (bar 1.5). fr1 6.2% to 8.0%, fr3 8.4% to 11.4%.
- H78 refuted, narrowly. OpenCV, 6 + 6, pooled: 31.6% against 27.0%, ×1.17 (bar 1.2). Left 28.5% to 32.8%, right 25.5% to 30.4%.
- H79 held. S0 false alarms of the chosen splits: left 1.40% / 1.32%, right 0.06% / 0.23%, fr1 0.35% / 0.71%, fr3 0.23% / 0.94% (6 + 6 / 5 + 6); the blind rates are 0.12–1.40%.
- H80 held. TUM pooled: ×1.63 at 5 + 6 (4.9% to 8.0%) against ×1.33 at 6 + 6.
- H81 held. The smallest twin sd of f<sub>new</sub> − f<sub>old</sub> picks a split that catches ±2% in 10.4% of TUM old sets at 6 + 6, 6.5% above π's 9.7%; it picks the same candidate as π in 39–50% of old sets at 6 + 6 and 17–18% at 5 + 6.
- Unscored. Ceiling (any candidate fires) at 6 + 6: 55.1%, 54.9%, 19.2%, 26.1% (left, right, fr1, fr3); the twin's pick reaches 60%, 55%, 42% and 44% of it. Within-old-set AUC at ±2%: 0.638–0.711 at 6 + 6 and 5 + 6. The twin's predicted gain (its mean π on the chosen over the blind mean) against the real one, 6 + 6 / 5 + 6: left 1.15 / 1.23 against 1.15 / 1.34, right 1.20 / 1.30 against 1.19 / 1.34, fr1 1.31 / 1.46 against 1.29 / 1.40, fr3 1.25 / 1.40 against 1.36 / 1.75. At 7 + 5: ×1.40 TUM, ×1.25 OpenCV. At 7 + 4 the choice fails on the Kinect: 4.9% blind to 2.5% chosen, where the blind rate is inflated by 3.4% false alarms (milestone 4j's over-alarm at 4 new views); OpenCV gains ×1.84 there. ±4% at 6 + 6: fr1 26.4% to 31.6%, fr3 38.2% to 49.0%.
- Reading. Milestone 4l's doubling on the TUM cameras came mostly from choosing the old set, which a robot cannot do. Holding the old calibration fixed and choosing among 7 candidate new sets gains a third (×1.33), and among 28 about two thirds (×1.63); more candidates help, so a robot that waits through more views should gain more, which 13 views per unit cannot show. The twin predicts the gain itself well (within 0.11 on 7 of 8 unit-designs at 5–6 new views) even though it overstates the level 1.2–1.9×: it is a sound way to compare candidates, not to state a detection rate. It still leaves most of the ceiling unused: some candidate catches the change in 19–55% of old sets, the twin's pick in 8–33%, so most of what separates a lucky new set from an unlucky one is noise in the corners the twin cannot foresee. The cheaper sd rule does as well, so no threshold draws are needed to choose. Choosing never raised false alarms past 1.4%, except where the test itself over-alarms (4 new views on the TUM cameras), and there choosing should not be used.
- Caveat. As set in advance: 13 views per unit, candidates that share views, old sets of 5–6 views (smaller than 4i–4l's 7), no interval, "waiting" modelled as best-of-all-candidates. A bigger view pool (the TUM sequences hold 820 and 916 qualifying frames) would test a real stopping rule; the TUM host refused this session's proxy (403), so that run waits for a session that can fetch it.
2026-11-15 · Milestone 4m planned: with the old calibration fixed, does waiting for the new view set the twin rates highest catch a 2% focal change more often?
Written before any milestone-4m quantity was computed. Milestone 4l's caveat: at 7 + 6 each old set leaves exactly one new set among a unit's 13 views, so choosing a split by the twin's predicted power π chose the old set as well. A robot holds one old calibration and can only choose which new views to test with. Here every old set is held fixed and its candidate new sets are the disjoint subsets of the views it leaves, so the choice is the robot's: same old set, pick the new set the twin rates highest. No new solve and no new data (the TUM host refused the proxy this session, so a larger view pool was not an option): milestone 4i's (OpenCV) and 4j's (TUM) cold-start solves and milestone 4l's predicted power, computed the same way (resid twin, 10,000 draws per unit, seed 20261114, σ̂ with the degrees-of-freedom correction) for more designs.
- Designs. 6 + 6: each of the 1,716 six-view old sets has 7 candidate six-view new sets (the 7 views it leaves, minus one). 5 + 6: each of the 1,287 five-view old sets has 28 candidates. 7 + 5 (6 candidates) and 7 + 4 (15) reported. An old set counts if at least 2 of its candidates are usable.
- Rules. Choose: the usable candidate with the highest π = (π<sub>0.98</sub> + π<sub>1.02</sub>)/2 (ties to the first in mask order). Random: the mean over the old set's usable candidates, the rate of picking one blind. The real detection is milestone 4i's per-split twin threshold on R, S1 at s, as scored in 4i, 4j and 4l. Rates are pooled over old sets and over s ∈ {0.98, 1.02}.
- H77 (waiting pays on the TUM cameras). 6 + 6, fr1 and fr3 pooled: the chosen rate is at least 1.5× the random rate. Refuted below.
- H78 (and on OpenCV's). 6 + 6, left and right pooled: at least 1.2×. Refuted below.
- H79 (without raising false alarms). S0 false alarms of the chosen splits at most 2.5% on each of the four units, at 6 + 6 and at 5 + 6. Refuted above on any.
- H80 (more candidates, more gain). On the TUM units pooled, the chosen / random ratio at 5 + 6 (28 candidates) exceeds the ratio at 6 + 6 (7). Refuted if not.
- H81 (no threshold draws needed). Choosing instead the candidate with the smallest twin sd of f<sub>new</sub> − f<sub>old</sub> (no reference quantile) gives a chosen rate, TUM pooled at 6 + 6, within 10% (relative) of π's. Refuted outside.
- Reported, unscored. The ceiling (the share of old sets where any candidate fires); the share of old sets where the chosen candidate fires and the blind mean is below one half; within-old-set AUC (pairs of candidates of the same old set); ±1% and ±4%; 7 + 5 and 7 + 4; π's predicted gain against the real one.
- Caveat set in advance. Thirteen views per unit, so the candidate pools are small (7 or 28) and share views with each other; the old sets at 6 + 6 and 5 + 6 are smaller than milestones 4i–4l's 7, so their own R is wider. Old sets share views, so no interval is given; two rigs, four units. "Waiting" is modelled as seeing every candidate and taking the best, not as a stopping rule over views that arrive in order.
2026-11-15 · Milestone 4l result: the twin ranks which updates can see a 2% focal change (AUC 0.66–0.75) and testing only its top quarter doubles the TUM cameras' power, 10.0% to 21.5%, at 0–2.3% false alarms; it overstates the level 1.2–1.9×
- Runs.
experiments/m4l_predict/analyse.py --rig opencv and --rig tum (about 5 min each on 4 cores): milestone 4i's twin draws (resid model, 10,000 per split, seed 20261114) rebuilt per unit; every per-split threshold reproduces the scored ones of milestones 4i and 4j (medians at 7 + 6: 10.93, 12.49, 9.52, 10.84; at 7 + 4 on fr3: 0.43, as 4j stored), and the real ±2% rates reproduce theirs (left 32.8%, right 33.4%, fr1 9.8%, fr3 10.3%). score.py scores H73–H76 and writes summary.json and power.png. tests/test_m4l.py checked first that the predicted power equals the normal-theory power on a Gaussian case, the AUC on known orderings, the quartile gate and σ̂'s degrees of freedom (N 1,404, p 237 on the 54-corner board). Before the scored run, a 500-draw check of the OpenCV rig (not kept) printed the per-design means, so the level on those two units was seen before scoring; nothing was changed after it. - H73 refuted, on OpenCV's cameras. AUC of π<sub>s</sub> for the real detection at ±2%, 7 + 6: left 0.659, right 0.693, fr1 0.749, fr3 0.748 (bar 0.70 on each).
- H74 held. Mean π over the real ±2% rate: left 1.88 (0.615 against 0.328), right 1.19 (0.399 against 0.334), fr1 1.34 (0.131 against 0.098), fr3 1.89 (0.194 against 0.103). The twin overstates power on all four units.
- H75 held. Top quarter of each TUM unit's splits by π: real ±2% rate 21.5% pooled (fr1 20.3%, fr3 22.6%), against 10.0% over all splits; the other three quarters: 6.3% and 6.2%.
- H76 held. S0 false alarms in the top quarter: left 0.70%, right 0.00%, fr1 2.35% (10 of 426; all splits 0.76%), fr3 0.00%.
- Unscored. OpenCV's top quarter: 48.3% (left) and 50.0% (right) against 32.8% and 33.4%. Absolute gates: π ≥ 0.3 passes 92.9% / 69.6% / 0% / 28.0% of splits (left / right / fr1 / fr3), with real ±2% rates 34.7%, 42.3%, none, 21.2%; π ≥ 0.5 passes 74.3% and 35.3% on OpenCV's (38.1%, 49.6%) and no TUM split. AUC at ±4%: 0.84, 0.88, 0.76, 0.86; at ±1%: 0.58, 0.59, 0.75, 0.71. Ranking by the twin's predicted sd of f<sub>new</sub> − f<sub>old</sub> alone: AUC 0.660, 0.690, 0.752, 0.743 (Spearman with π 0.98 on each unit), so the cheaper gate does as well. σ̂ without the degrees-of-freedom correction: level ratios 2.09, 1.40, 1.63, 2.26. Every split that holds one of milestone 4j's far solves (8 on fr1, 36 on fr3) sits in the bottom quarter by π. At 7 + 4 the prediction fails on the TUM units: AUC 0.42 (fr1) and 0.64 (fr3), level ratio 0.75 and 0.41, and fr3's top quarter raises 5.0% false alarms; at 7 + 5 AUC 0.64–0.73 and ratios 1.48–1.82.
- Reading. A unit's own twin can say before a test runs which updates are worth testing. On the hand-held TUM captures, testing only the quarter of updates the twin rates highest doubles the chance of catching a 2% focal change, and every split that carries a far, near-degenerate solve is in the quarter it rates lowest; the sd of the twin's f<sub>new</sub> − f<sub>old</sub> carries the same ranking, so a robot needs no threshold draws to make the choice. The ranking is moderate (AUC 0.66–0.75), weakest on OpenCV's cameras where most splits are already informative, and the twin's absolute power is 1.2–1.9× too high: the real f's vary more than the linear twin fed with the unit's residual patterns says, so an absolute gate ("test when π ≥ 0.5") is not safe to read as a power figure. Use the twin to rank or to wait for a better view set, not to promise a detection rate. At 4 new views the twin's delete-one sets are 3-view and the prediction breaks down, a third reason to collect at least 6.
- Caveat. At 7 + 6 each old set leaves exactly one new set among the 13 views, so choosing a split by π chooses the partition, old set included; a robot holds one old set and can only wait for better new views, and that gain is not measured here (it needs more than 13 views per unit). The quarters share views, so no interval is given; two rigs, four units. fr1's top-quarter false alarms (2.35%) sit three times its all-split rate and just inside the bar.
2026-11-15 · Milestone 4l planned: can the unit's twin tell, before the test runs, which updates are able to catch a 2% focal change?
Written before any milestone-4l quantity was computed. Milestones 4j and 4k found that power is a property of the unit's views: a 2% focal change at 7 + 6 views is caught in 33% of splits on OpenCV's cameras and 10% on the TUM ones, and the far, near-degenerate subset solves are real minima, not a solver artefact. A fleet cannot change how informative a camera's views are, but it can decide which updates to run the test on, or wait for more views. The twin of milestone 4i already holds what that decision needs: for each (old, new) split, its draws give the reference threshold, and the same draws with a shift added give the twin's own predicted power. Here that prediction is scored against the real detections, split by split, on all four units.
- Predicted power. For each 7 + 6 split and s in {0.98, 1.02}, π<sub>s</sub> = the share of the split's 10,000 twin draws (resid model, milestone 4i's seed and chunks, so the thresholds are the scored ones of 4i and 4j) with z<sub>R</sub><sup>2</sup> = (σ̂·Δ + (s − 1)·f<sub>13</sub>)<sup>2</sup> / (σ̂<sup>2</sup>(R<sub>old</sub> + s<sup>2</sup>R<sub>new</sub>)) above the split's own twin threshold, where Δ and R are the twin's unit-scale draws and σ̂ = rms·√(N/(N − p)) is the unit's 13-view residual scale with the degrees-of-freedom correction (N residuals, p free parameters). π = (π<sub>0.98</sub> + π<sub>1.02</sub>)/2. Nothing is fitted to the real detections. Code:
experiments/m4l_predict/analyse.py, which imports milestone 4i's (left, right) and 4j's (fr1, fr3) analysis unchanged; real statistics are their cold-start solves. - H73 (the twin ranks the splits). AUC of π<sub>s</sub> for the real detection (R against its per-split twin threshold, S1 at s), pooled over the two signs: ≥ 0.70 on each of the four units. Refuted below on any.
- H74 (and gets the level). Mean π over the splits within a factor of 2 of the real ±2% rate on each unit (0.5 ≤ ratio ≤ 2). Refuted outside on any.
- H75 (choosing the update doubles the power on the TUM units). In the top quartile of each TUM unit's splits by π, the real ±2% rate pooled over the two units and signs is ≥ 20% (all splits: 10.0%). Refuted below.
- H76 (without raising false alarms). S0 false alarms in the same top quartile ≤ 2.5% on each of the four units. Refuted above on any.
- Reported, unscored. The OpenCV units' top quartile; absolute gates (π ≥ 0.3, 0.5) with the share of splits that pass; σ̂ without the correction; 7 + 4 and 7 + 5; ±1% and ±4%; the ranking by the twin's predicted sd of f<sub>new</sub> − f<sub>old</sub> alone (no threshold), as a cheaper gate; whether the real far solves of milestone 4j sit in the bottom quartile.
- Caveat set in advance. Selecting splits by π selects old sets as well as new ones; a robot has one old set, so the gate's practical form is "wait for a new set the twin says can see the change". The splits reuse views, so quartiles share views and no interval is given; two rigs, four units.
2026-11-15 · Milestone 4k result: the far solves on the TUM cameras are real minima of near-degenerate view sets, not a cold-start artefact; a warm start reproduces every one of them, and a basin check that drops them raises the false alarms at 4 new views on the Xtion from 4.7% to 20%
- Runs. 11,442 warm-started solves (
experiments/m4k_warm/run.py, f started at f<sub>P</sub> = 527.1 px, about 85 min on 4 cores); milestone 4j's analysis run unchanged on arms W, WB and CB (analyse.py, about 10 min); compare.py sets each subset's warm and cold solves side by side (post hoc). tests/test_m4k.py checked first that f<sub>P</sub> is TUM's mean, that the basin check drops only solves outside ±10% and leaves its input untouched, and that a warm and a cold start land on the same f on a well-conditioned 6-view set. - H68 refuted. Arm W, 6-view solves more than 5% from the median: 76 (fr1) and 30 (fr3), against 8 and 30 cold. On fr3 the warm start changes nothing: every 5-, 6- and 7-view subset that converged from both starts lands on the same f (1,285 / 1,285, 1,716 / 1,716, 1,716 / 1,716). On fr1 it is worse: the 13-view solve from f<sub>P</sub> stops at f = 508.88 px with rms 0.483 px, against the cold start's 526.47 px at 0.152 px (cost 145.5 against 14.5), and the cold start reaches a lower cost on 89 of the 6-view and 98 of the 7-view sets, the warm start on none.
- H69 refuted. Arm W, 3-view sets converged: 96.5% (fr1, 276 / 286) and 80.8% (fr3, 231 / 286), against 95.8% and 82.9% cold.
- H70 refuted. Arm WB, S0 at k = 4: 4.36% (fr1) and 20.09% (fr3), against 3.27% and 4.73% cold.
- H71 refuted, on fr1 at k = 6. Arm WB: k = 5 1.28% / 0.88%, k = 6 2.13% / 0.18% (fr1 / fr3).
- H72 refuted. Arm WB, ±2% at k = 6 pooled: 10.6% (cold 10.0%).
- Unscored. Arm CB (cold solves with the basin check, which leaves the fr1 13-view solve and twin as in milestone 4j): S0 8.09% / 19.09% at k = 4, 1.02% / 0.88% at 5, 0.71% / 0.18% at 6; ±2% at k = 6 10.5%, ±4% 43.1%, ±1% 2.1%. The basin check leaves 275 of fr1's 715 and 89 of fr3's 715 four-view sets usable (cold: 686 and 587), because most 3-view solves lie more than 10% from f<sub>P</sub> (fr3: 86 of 286 survive). Arm W: S0 2.12% / 4.66% at k = 4, 3.35% / 0.69% at 5, 5.28% / 0.06% at 6; ±2% 11.0%. S3 at k = 6 (WB): 5.4% / 8.3%. Harmful adoptions (R, per-split twin): cold 0.29%, W 0.46%, CB 0.73%, WB 1.12%.
- Post hoc, computed after the scored run. All 38 of milestone 4j's far 6-view cold solves (8 on fr1, including f = 9,289 px, and 30 on fr3) come back from the warm start at the same f to 10<sup>−4</sup>. fr3's 30 fit their corners as well as the 13-view solve does (median rms 0.071 px against 0.075 px): they are minima of the data, in view sets whose geometry leaves f nearly free, and the cost surface has no lower point nearer f<sub>P</sub>.
- Reading. Milestone 4j's caveat that "the far solves could be removed by a warm start or a basin check" was wrong on both counts. A warm start cannot remove a far solve that is the subset's true minimum, and on these two cameras every one of them is. Starting f at the spec does not start the rest of the problem well either: the poses and board come from a fronto-parallel guess at that f with the principal point at the image centre and k1 = 0, and on the strongly distorted Kinect that start falls into a worse minimum on the full 13 views and on 5–6% of 6–7-view sets (why the 1.3× start avoids it was not studied). A basin check that drops solves far from the spec does harm: it removes exactly the delete-one values that widen the robust jackknife when a view set is near-degenerate, so the surviving sets' R is too narrow for their twin threshold, and false alarms at 4 new views rise to 8–20%. The power gap on the TUM cameras (10% against 33% on OpenCV's) is not the solver's doing: it is how little these hand-held views pin f down. For a fleet: keep the lowest-cost solve over several starts (here the cold start is never beaten on a set of 6 or more views), never clip f to a window, and treat a wide delete-one spread as information the test needs.
- Caveat. Two units, one sequence each; the splits reuse views, so no interval is given. One warm start was tried (f only); a start that also sets the principal point and k1 from the spec, or poses from a previous solve, was not.
2026-11-15 · Milestone 4k planned: warm-start every subset solve from the fleet's nominal focal length, with a basin check, and re-score the TUM units
Written before any milestone-4k solve was run. Milestone 4j found that on the two TUM RGB-D cameras the cold-start solver (start f at 1.3× the k1 Zhang f: 677.0 and 689.8 px, 28–29% above the units' own 13-view f) lands in far minima on some subsets (6-view solves more than 5% from the median: 8 of 1,709 on fr1, up to f = 9,289 px, and 30 of 1,716 on fr3; none on OpenCV's cameras) and fails to converge on 4% and 17% of 3-view sets. A robot in a fleet never starts cold: it knows the spec of its camera model. Here every subset is re-solved from that spec, and a basin check fixed now drops any solve the spec says is implausible.
- Start. f<sub>P</sub> = the mean over the two units of TUM's published (f<sub>x</sub> + f<sub>y</sub>)/2 = (516.9 + 537.3)/2 = 527.1 px, one figure for both units, as a fleet with one spec sheet would hold; principal point at the image centre and k1 = 0 as before. Nothing else in milestone 4f's solver changes (
experiments/m4k_warm/run.py): every subset of 3–7 views and all 13, both units, 11,442 solves. The twin is refitted from the same start and must reproduce the new 13-view f. - Arms. W: warm start alone. WB (scored): warm start, and a solve with |f/f<sub>P</sub> − 1| > 10% counts as not converged (a far solve then cannot enter as a set or a delete-one value; 10% is wider than any change the test is meant to catch, and the units' published f are within 2% of f<sub>P</sub>). Milestone 4j's splits, scenarios, statistics and per-split twin references (resid model, 10,000 draws, seed 20261114) are run unchanged on each arm by importing
m4j_tum/analyse.py. - H68 (the warm start stays in the basin). Arm W: 6-view solves more than 5% from the unit's median 6-view f number at most 2 on each unit. Refuted above on either.
- H69 (and converges on small sets). Arm W: ≥ 97% of 3-view sets converge on each unit (cold: 95.8% and 82.9%). Refuted below on either.
- H70 (false alarms at 4 new views come down). Arm WB, R against its per-split twin threshold, S0 at k = 4: ≤ 2.5% on each unit (cold: 3.27% and 4.73%). Refuted above on either.
- H71 (and stay low at 5–6). Arm WB, the same at k = 5 and at k = 6: ≤ 1.5% on each unit and each k (cold: 0.06–0.80%). Refuted above on any.
- H72 (power recovers part of the way). Arm WB, ±2% at k = 6 pooled over sign and unit: ≥ 15% (cold: 10.0%; OpenCV's cameras: 33.1%). Refuted below.
- Reported, unscored. Arm W's rates; the cold solves with the same basin check (CB; milestone 4j's solves were already looked at, so this arm cannot be scored); the 13-view f and its change from cold; S3 (the unit swap); ±1% and ±4%; harmful adoptions; far solves at 4, 5 and 7 views; solver evaluations per solve.
- Caveat set in advance. The scaled-focal scenarios scale the new set's solved f after the solve, as in milestones 4g–4j, so the warm start sits at the unscaled spec; a real 2% change would start 2% from f<sub>P</sub>, well inside the 10% check. Two units, one sequence each; the splits reuse views, so no interval is given.
2026-11-14 · Milestone 4j result: on two more cameras (TUM RGB-D's Kinect and Xtion) the twin reference stays quiet at 5–6 new views (0.1–0.8% false alarms) but over-alarms at 4 (3.3% and 4.7%), and a 2% focal change is caught in 10% of 6-view updates, not 33%
- Runs. 13 views per unit by
prepare.py's rule (fr1: 820 of 1,507 frames qualified; fr3: 916 of 1,327); 11,442 solves (run.py, 2 h on 4 cores, most of it fr1, whose 3-view sets often run to the 2,000-evaluation cap). Converged: fr1 274 / 286 three-view sets, 698 / 715 four-view, ≥ 99.5% at 5–7; fr3 237 / 286, 708 / 715, ≥ 99.8%. Usable 4-view sets (≥ 3 of their 4 delete-one solves converged): 686 and 587 of 715. analyse.py (milestone 4g's splits and 4i's twin, imported unchanged, about 5 min); check.py 1,280 cold-start solves (8 min). tests/test_m4j.py checked first that the corners load, that the 8×6 board is wired through the solver, and that the twin on 48 corners matches the solver. - H63 held. fr3's 13-view f is 533.87 px against TUM's (fx + fy)/2 = 537.3: −0.64%. fr1 (unscored): 526.47 against 516.9, +1.85%, the bias expected from fitting k1 alone to a lens TUM describes with d1 = −0.95. Rms 0.152 px (fr1) and 0.075 px (fr3).
- H64 refuted, on fr1. fr3: median sd(nonlinear Δf)/sd(g<sub>S</sub>·ε) 1.000–1.004 for n = 4–7, correlation ≥ 0.993. fr1: 1.002 at n = 5 and 6, but at n = 4 one of the four random sets never converged from the cold start (40 of 40 draws; the other three read 1.01–1.08), so the median is undefined; and at n = 7 two of four sets jumped to far minima (24 of 160 draws more than 5 sd from the linear prediction; sd ratio up to 9.6, median 2.61). The linear prediction holds where the solver stays in its basin; the misses are the cold-start solver's.
- H65 refuted, at 4 new views. R against its per-split twin threshold, S0: fr1 3.27% at k = 4, 0.80% at 5, 0.76% at 6; fr3 4.73%, 0.62%, 0.06% (k = 4–6 pooled 2.47% and 3.48%; bar 0.5–2.5%). The 4-view sets' delete-one solves are 3-view, where the twin is least trustworthy (milestone 4i) and the real solves fail most often.
- H66 refuted. ±2% at k = 6 pooled over sign and unit: 10.0% (fr1 11.0 / 8.6%, fr3 10.1 / 10.4%; milestone 4i: 33.1%). ±4%: 33.7–48.0%; ±1%: 1.5–2.8%.
- H67 refuted. The units' own 13-view f differ by 1.40%, not TUM's 3.5%: fr1's k1-only bias (+1.85%) closes most of the gap. S3 at k = 6: R fires on 6.3% (fr1's new views against fr3's old) and 7.9% (the reverse). That sits between the same test's ±1% (2.0%) and ±2% (10.0%) rates on scaled focal lengths, so the real change is caught about as often as a scaled one of its size.
- Unscored. Per-split twin thresholds at 7 + 6, 10th / 50th / 90th percentile: 6.88 / 9.52 / 17.33 (fr1) and 8.62 / 10.84 / 16.60 (fr3), near milestone 4i's 11–13. At k = 6 the plain jackknife against its twin reference catches ±2% in 12.1% (false alarms 0.82% / 0.12%), one threshold per design 8.0%, R against χ²<sub>1</sub> 16.7% at 1.23% / 0.52%, R against milestone 4h's 18.45 3.2%, J against χ²<sub>1</sub> 1.8%. The iid noise model: ±2% 9.3%, S0 over k = 4–6 2.92% / 5.67%. Harmful adoptions: 0.29% of decisions (R, per-split twin), 0.18% (R, χ²<sub>1</sub>).
- Post hoc, computed after the scored run. Why the power fell: the robust spread of f over all 6-view sets is 0.73% of f (fr1) and 0.51% (fr3), against 0.51% and 0.30% on OpenCV's cameras (milestone 4f's solves), and the cold-start solver lands far away on some subsets that OpenCV's views never produced: 6-view solves more than 5% from the median, 8 of 1,709 (fr1, up to f = 9,289 px) and 30 of 1,716 (fr3), against 0 of 1,716 on each OpenCV camera. A far solve that enters as a delete-one value inflates J (J against χ²<sub>1</sub> barely fires) and, as a new set, fires the test whatever the truth. The real-spread-to-predicted ratio milestone 4g used is dominated by those solves here (R at 7 + 6: 10.4 on fr1) and is not reported as a finding.
- Reading. The twin reference transfers to two different cameras, lenses and boards in the one way that matters most for a fleet: at 5 and 6 new views it stays at or below its nominal 1% (0.06–0.80%) with nothing tuned on these units. It does not transfer at 4 new views, where its delete-one sets are 3-view and both the twin and the real solver leave their linear regime; the paper's recommendation to collect at least 6 new views before testing now has a second reason. The 33% power does not transfer: power is a property of the unit, set by how precisely 6 views fix its focal length and by how often the cold-start solver falls out of its basin, and these hand-held 640×480 captures give 10%. A real 1.4% difference between two units is caught as often as a scaled change of that size, which is the first check of the test on a change no script applied.
- Caveat. Two units, one sequence each (about 50 s, one person holding the board), 13 views; the splits reuse views, so no interval is given. fr1's lens is not described by k1 alone. The far solves could be removed by a warm start or a basin check, which would change the power figures; that is a separate, pre-registered step.
2026-11-14 · Milestone 4j planned: the twin reference and the update test on two more units (TUM RGB-D's Kinect and Xtion)
Written before any milestone-4j focal length was looked at. Every result of milestones 1–4i comes from one stereo rig (OpenCV's two cameras, one board, 13 views), and the paper's Section 10 asks first for the twin reference and the robust test "pre-registered on a second unit's views". The TUM RGB-D benchmark (Sturm et al., IROS 2012; CC BY 4.0) publishes checkerboard calibration sequences for each of its cameras, with the authors' own calibration: freiburg1_rgb_calibration (a Microsoft Kinect, fx 517.3, fy 516.5, strong distortion d0 0.2624, d1 −0.9531) and freiburg3_calibration_rgb_depth (an Asus Xtion whose images come undistorted, fx 535.4, fy 539.2). Two new units, a different board (8×6 inner corners), a different sensor and lens, and a real focal difference between them (3.5% in TUM's figures), which gives the swap scenario a true change instead of a scaled one.
- Data. 13 views per unit picked by a rule written into
experiments/m4j_tum/prepare.py before any solve: full-board detections with 5×5 corners, the even board's 180° ambiguity resolved by the first row pointing right, corner spacing ≥ 15 px, then 13 equal-count blocks in time and the most foreshortened of each block's sharper half. A first version of the rule (sharpest frame per time block) picked boards 7–8 px between corners and 11 of 13 fronto-parallel views; it was replaced after looking only at corner spacing and a contact sheet, no solve. Only corners and frame names are stored (corners_fr1.json, corners_fr3.json); a person holds the board in every frame, so no image is committed (fetch.sh gets the sequences). fr1: 820 of 1,507 frames qualify; fr3: 916 of 1,327. - Solves. Milestone 4f's solver unchanged (k1 only, start 1.3× the k1 Zhang f, references 0 and n // 2): every subset of 3–7 views and all 13, both units, 11,442 solves (
run.py). Splits, scenarios, J and R exactly as milestone 4g, with the two TUM units in place of left and right: S0, S1 (s = 0.96–1.04) and S3 (the old set from the other unit). The twin and its references exactly as milestone 4i (resid model scored, iid reported, 10,000 draws per split, seed 20261114). - H63 (anchor). fr3's 13-view f lies within 1% of TUM's (fx + fy)/2 = 537.3 px. Refuted outside. (fr1's is reported against 516.9 only: the k1 model cannot follow its d1 = −0.95, so a bias is expected there.)
- H64 (the linearisation holds on these units). As H59 with 4 random sets of each size 4–7 per unit, 40 draws each, iid noise at the unit's 13-view rms: the median over sets of sd(nonlinear Δf)/sd(g<sub>S</sub>·ε) lies in 0.8–1.25 for each n on both units. Refuted outside.
- H65 (the twin reference is near nominal on new units). R against its per-split twin threshold, S0 false alarms over k = 4–6 in 0.5–2.5% on each unit (milestone 4i: 1.44% and 0.37%). Refuted outside on either.
- H66 (power carries over). ±2% at k = 6, pooled over sign and unit, detected in ≥ 25% of splits (milestone 4i: 33.1%). Refuted below.
- H67 (a real unit swap is caught). S3 at k = 6 (old 7 views from the other unit, new 6 from this one; the old set's reference drawn from the other unit's twin, as in milestone 4i): R against its per-split twin threshold fires on ≥ 50% of splits on each unit. Refuted below on either.
- Reported, unscored. fr1's f against TUM's; both units' 13-view rms; J against its twin reference; one twin threshold per design; R against χ²<sub>1</sub> and against milestone 4h's independent-offset reference (18.45 at 7 + 6; unit-free); the iid model; ±1% and ±4%; k = 4 and 5; harmful adoptions; the twin's and the real scale ratios (milestone 4i's post hoc check).
- Caveat set in advance. fr1's lens is not described by k1 alone, so its residuals carry the unmodelled distortion; the resid noise model passes that pattern into the twin, which is the point of a per-unit twin, but a large model error could also bend H64. The 13 views of each unit come from one hand-held sequence of about 50 s, so they are not independent sessions; no interval is given.
2026-11-14 · Milestone 4i result: a reference from the unit's own linearised twin catches a 2% focal change from 6 views in 33% of splits, not 19%, at 0.4–1.4% false alarms; once the reference knows the geometry, the plain jackknife does as well as the robust one
- Runs. The twin is each camera's 13-view 5×5 solve, re-run to keep its poses and points (f 534.336 and 535.753 px, equal to milestone 4f's stored solves to 10<sup>−8</sup>; rms 0.114 and 0.153 px). Gains g<sub>S</sub> for every subset of 3–7 views (5,720 per camera, about 15 s), 10,000 noise draws per split, every split of milestone 4g at 7 + 4, 7 + 5 and 7 + 6 (
experiments/m4i_geomref/analyse.py, about 4 min). tests/test_m4i.py checked first that the twin reproduces the stored solve, that a gain matches a small nonlinear perturbation, that it does not depend on which view fixes the gauge, and that a per-split reference fires on 0.6–1.4% of independent twin draws. check.py ran 9,600 cold-start nonlinear solves on twin corners (25 min on 4 cores). - H59 held. sd(nonlinear Δf)/sd(g<sub>S</sub>·ε), median over sets, n = 4–7: 1.0001–1.0004 (left) and 0.9990–1.0017 (right); median correlation ≥ 0.9995. At 3× the real rms: 0.999–1.018. The resid model: 0.9995–1.005. Not at 3 views: some 3-view sets leave the linear regime (left, iid: 40 of 320 draws more than 5 sd from the linear prediction; resid: 52 left, 24 right; one right set's nonlinear sd is 23× the linear one). 3-view sets enter only as the delete-one sets of 4-view new sets, so the twin is least trustworthy at 7 + 4.
- References. resid model, 7 + 6: the per-split 99th percentile of z<sub>R</sub>² has median 10.93 (left) and 12.49 (right), 10th–90th percentile 8.78–14.69 and 9.81–16.15; one percentile pooled over the design: 11.64 and 12.90. Milestone 4h's independent-offset model gave 18.45, χ²<sub>1</sub> 6.63. Medians at 7 + 4: 8.73 and 7.38; at 7 + 5: 11.85 and 12.57. For the plain jackknife J: median 3.83 and 4.13 at 7 + 6.
- H60 refuted, on the right camera. R against its per-split twin threshold, S0 over k = 4–6: 1.44% (left, inside 0.5–2.5%) and 0.37% (right, below).
- H61 held. ±2% at k = 6 pooled over sign and camera: 33.1% (per camera and sign 31.5–35.3%), against 18.6% for milestone 4h's model reference.
- H62 held. 90th over 10th percentile of the per-split threshold at 7 + 6: 1.67 (left) and 1.65 (right).
- Unscored. ±4% at k = 6: 81.2–84.7% (model reference: 66.4–75.1%); ±1%: 4.1–10.5%; ±2% at k = 5: 18.2%, at k = 4: 8.9%. The plain jackknife J against its own per-split twin threshold: ±2% at k = 6 in 34.7%, false alarms 1.59% / 0.35%. One twin threshold per design instead of per split: R 30.8% at 1.60% / 0.66%. The iid noise model instead of the unit's residual patterns: R 30.7% at 1.74% / 0.42%. The swap stays invisible to f alone (0.5% at k = 6). Harmful adoptions: 0.24% of decisions (per-split), 0.36% (per-design). Each split's threshold correlates only weakly with its real z<sub>R</sub>² under S0 (Spearman 0.24 left, 0.10 right).
- Post hoc, computed after the scored run. Under the twin (resid), sd(f<sub>new</sub> − f<sub>old</sub>) over the root mean predicted variance at 7 + 6 is 0.75 / 0.76 for R and 0.52 / 0.54 for J; on the real views milestone 4g measured 0.88 / 0.48 and 0.57 / 0.31. At 7 + 4 the twin gives R 0.31 / 0.29 against the real 0.31 / 0.25. So the twin reproduces most of the jackknife's small-set overstatement on the left camera from the views' geometry alone, and the right camera's jackknife overstates its real spread about 1.6× more than its twin says. That is why the right camera sits below the bar: the same extra the independent-offset model missed in milestone 4f (its views do not add independent offsets).
- Reading. The jackknife's overstatement at 4–7 views is mostly geometry: a linear twin built from the unit's own solve predicts it, and a reference drawn from that twin is 11–13 at 7 + 6 views rather than the independent-offset model's 18.45. Against it the robust test fires on 1.4% and 0.4% of unchanged updates and catches a 2% focal change from 6 views in a third of splits, 1.8× milestone 4h's honest figure, with nothing borrowed and from solves the robot already ran. It does not reach the 43% of the borrowed threshold or the 54% of χ²<sub>1</sub>, whose 2–3% false alarms rested on cancelling errors. Two side findings change the paper's recommendation. First, with a geometry-aware reference the MAD buys nothing: the plain jackknife catches 34.7%. Second, one threshold per design does almost as well as one per split (30.8%); the per-split spread is real (H62) but barely tracks the real statistics, so most of the gain is the twin getting the design's tail right.
- Caveat. One rig, one board, 13 shared views; no interval is given. The twin is linear and is fed the unit's own 13 residual patterns, so error sources the residuals do not show (the right camera's extra spread) leave the test conservative there rather than lenient. That is the safe direction for a fleet, but it means the 0.37% is a property of this unit, not a guarantee.
2026-11-14 · Milestone 4i planned: a reference for each split from the unit's own linearised bundle adjustment
Written before any milestone-4i statistic was computed on the real splits. Milestone 4h gave the robust jackknife test (R) a reference with nothing borrowed, from a model in which each view adds an independent offset to f. On this rig it was conservative (0.2–0.5% false alarms) and weak (±2% at 6 new views caught in 19% of splits), because the model ignores the views' geometry: a small set near the critical configuration has delete-one solves that swing far (milestone 4f), which widens R in a way independent offsets do not. A reference that knows the geometry should fire near its nominal 1% and recover part of the power χ²<sub>1</sub> seemed to have (54%). Here the robot's own 13-view solve is the twin: for any view subset S its first-order change in f under a pixel perturbation ε is g<sub>S</sub>·ε, with g<sub>S</sub> = J<sub>S</sub>(J<sub>S</sub><sup>T</sup>J<sub>S</sub>)<sup>−1</sup>e<sub>f</sub> from the subset problem's Jacobian at the twin (experiments/m4i_geomref/twin.py). One matrix product then draws every set's f and its delete-one f's, and so z<sub>R</sub><sup>2</sup>, for each (old, new) split, and the split's own 99th percentile is its threshold. A unit can compute it from solves it already ran; nothing is borrowed from another unit.
- Data. Milestone 4g's splits, scenarios and real statistics, imported unchanged; the twin is the 13-view 5×5 solve of milestone 4f (re-run to keep its poses and points; it must reproduce the stored f to 10<sup>−8</sup>). No new real solve.
- Noise models (the statistic is scale-free, so only the pattern matters). resid (scored): each view gets the 13-view fit's residual pattern of a view drawn at random from the unit's own 13, with a random sign, so the within-view correlation milestone 6 found (columns, the bent board) is carried to every view. iid: independent N(0, 1) per pixel coordinate (reported). 10,000 draws per split, seed 20261114; designs 7 + 4, 7 + 5, 7 + 6; S3 (swap) uses the other camera's twin for the old set.
- H59 (the linearisation holds for this solver). Synthetic corners from each camera's twin plus iid noise at the real 13-view rms, solved by the same cold-start solver as the real runs (8 random sets of each size 3–7, 40 draws each,
check.py): the median over sets of sd(nonlinear Δf)/sd(g<sub>S</sub>·ε) lies in 0.8–1.25 for each of n = 4, 5, 6, 7 on both cameras. Refuted if any falls outside. (3× the rms and the resid model reported.) - H60 (the per-split twin reference is near nominal on real views). R against its per-split threshold (resid model), S0 false alarms over k = 4–6 in 0.5–2.5% on each camera. Refuted outside on either.
- H61 (and recovers power). ±2% at k = 6, pooled over sign and camera, detected in ≥ 25% of splits (milestone 4h's model reference: 18.6%). Refuted below.
- H62 (geometry matters split by split). Over the 1,716 splits at 7 + 6, the ratio of the 90th to the 10th percentile of the per-split threshold is ≥ 1.5 on both cameras. Refuted below on either (then one threshold per design would do).
- Reported, unscored. The iid model; a per-design twin reference (one 99th percentile pooled over the design's splits); J against its own twin references; ±1% and ±4%; k = 4 and 5; the swap; harmful adoptions; the Spearman correlation of each split's threshold with its real z<sub>R</sub><sup>2</sup> under S0.
- Caveat set in advance. The twin is linear and its noise comes from 13 residual patterns of one rig, so a unit whose errors are dominated by something the residuals do not show (a shared offset across views, for instance) would not be described by it. The 4f solves showed that the real spread of f is 2–6× the solver's σ, which a scale-free statistic absorbs only if the extra spread has the same pattern over views as the pixel noise. One rig, one board, 13 shared views: no interval is given.
2026-11-14 · Milestone 4h result: a reference derived from the independent-offset model keeps the robust jackknife test at 0.2–0.5% false alarms, but it catches a 2% focal change from 6 views in 19% of splits, not 54%; χ²<sub>1</sub> looked right only because two errors cancel
- Runs. No new solve: milestone 4g's splits, scenarios and statistics, imported unchanged. Model references by Monte Carlo, 10<sup>6</sup> unchanged updates per design and offset law (seed 20261114), 400,000 per set size for n<sub>old</sub> = k = 4..40 (
experiments/m4h_modelref/analyse.py, 30 s). tests/test_m4h.py checked first that the vectorised variances equal milestone 4g's, that the plain jackknife's reference is a Welch-t one (99th percentile of z² 9.0–11.0 at 7 + 6), that a reference from one seed fires on 0.8–1.2% of another seed's draws, and that the statistic is scale-free. - References. Gaussian offsets, 99th percentile of z<sub>R</sub>²: 25.84, 22.90 and 18.45 at 7 + 4, 7 + 5 and 7 + 6 views (9 + 4: 23.28), against χ²<sub>1</sub>'s 6.63; of z<sub>J</sub>²: 11.93, 10.44, 9.74. Student-t offsets make R's tail heavier still (7 + 6: 21.56 with t<sub>5</sub>, 31.71 with t<sub>3</sub>) and J's lighter (8.97, 8.39). Under the model, R against χ²<sub>1</sub> fires on 8.5%, 7.2% and 6.1% of unchanged updates at k = 4, 5, 6.
- H56 held. R against its Gaussian model reference, S0 over k = 4–6: 0.45% (left) and 0.20% (right) false alarms.
- H57 held. R against the model reference, ±2% at k = 6 pooled over sign and camera: 18.6% (per camera and sign 17.7–20.0%), against 53.8% for R against χ²<sub>1</sub>.
- H58 held. Under the Gaussian model with n<sub>old</sub> = k = n, P(z<sub>R</sub>² > 6.63) is 10.7% at n = 4, 3.8% at 10, 2.7% at 15 and 2.2% at 20; it reaches 1.6% only at 40. The model's 99th percentile is 12.0 at n = 10 and 8.84 at 20.
- The cancellation. On the real views, R against χ²<sub>1</sub> fired on 3.0%, 4.6% and 3.1% (left) and 1.1%, 1.4% and 0.2% (right) of unchanged updates at k = 4, 5, 6, below the model's 8.5%, 7.2% and 6.1%. The robust jackknife still overstates the real spread at these set sizes (milestone 4g, H53: real/predicted 0.88 and 0.48), and that overstatement offsets χ²<sub>1</sub>'s leniency. Milestone 4f found the overstatement fades as sets grow; the model's leniency fades more slowly (still 2.2% at 20 + 20), so on a unit with more views R against χ²<sub>1</sub> should drift towards the model's rates, not stay at 2–3%.
- Unscored. At k = 6 the model reference catches ±4% in 66.4–75.1% of splits (χ²<sub>1</sub>: 92.2–94.3%) and ±1% in 2.2–4.2%; at k = 5, ±2% in 6.0–10.8%. With the same calibrated basis the plain jackknife against its own model reference catches ±2% at k = 6 in only 8.6% pooled, so the robust scale does buy power (18.6% against 8.6%) once both tests have a derived reference. The t<sub>5</sub> and t<sub>3</sub> references give 14.8% and 7.8% at ±2%, k = 6. The unit swap is invisible to f alone under every reference (0.3% at k = 6). Harmful adoptions over all designs and scenarios: 0.63% (R, χ²<sub>1</sub>), 0.14% (R, model reference).
- Reading. Milestone 4g's 54% is not the power of a 1% test. It is the power of a test whose reference is about 6× too lenient in theory and happens to land near 2–3% on this rig because the jackknife is still too wide at 6 and 7 views. A reference derived from the independent-offset model is the only one in this programme that borrows nothing and rests on a stated assumption, and with it the f-alone test is quiet (0.2–0.5%) but catches a 2% change in only 19% of 6-view updates. Between that and the borrowed threshold's 43% (milestone 4g) lies the price of borrowing nothing at 6 new views. For a fleet the practical choice is (a) the model reference, conservative on this rig, which a robot can compute for any set sizes in about a second, or (b) χ²<sub>1</sub> as an approximate 2–5% test, with its real rate checked on each unit. The paper's recommendation 2 has to change accordingly.
- Caveat. One rig, one board, 13 shared views; no interval is given. The model ignores the near-critical behaviour of small sets that milestone 4f found, which is why it is conservative here; a unit whose views add independent offsets would sit at the nominal 1%.
2026-11-14 · Milestone 4h planned: a reference for the robust jackknife that comes from a model, not from χ²<sub>1</sub> or a sibling
Written before any milestone-4h statistic was computed on the real solves. Milestone 4g's best test (the robust, MAD-based jackknife R against χ²<sub>1</sub> at 1%) caught a 2% focal change in 54% of 6-view updates at 2.3% false alarms. But χ²<sub>1</sub> was never derived for R: a MAD of 6 or 7 delete-one values is itself noisy, so z<sub>R</sub>² should have a heavier tail than χ²<sub>1</sub>. A first Monte Carlo of the ideal case (each view adds an independent Gaussian offset to f, R and z<sub>R</sub>² formed exactly as m4g_allsplits/analyse.py does; 400,000 draws, seed 0; no real solve read) gives a 99th percentile of z<sub>R</sub>² of 18.4 at 7 + 6 views and 22.6 / 25.8 at 7 + 5 / 7 + 4, and says R against χ²<sub>1</sub> would fire on 6.0–8.5% of unchanged updates. If so, milestone 4g's 2.3% on the real views is two errors cancelling: a lenient reference against a robust jackknife that still overstates the real spread (milestone 4g, H53: real/predicted 0.88 and 0.48). Nothing in that cancellation would survive a unit with more views, where the overstatement fades (milestone 4f). Here the test gets a reference from the model itself, which a robot can compute for any set sizes with nothing borrowed, and the model's own verdict on when χ²<sub>1</sub> becomes safe.
- Data. Milestone 4f's stored 5×5 solves and milestone 4g's splits, scenarios and statistics, unchanged (
experiments/m4h_modelref/analyse.py imports them). No new solve. - Model references. For each design (n<sub>old</sub>, k), q<sub>M</sub> = the 99th percentile of z<sub>R</sub>² (and of z<sub>J</sub>² for J) when every view's offset is drawn iid from N(0, 1), by Monte Carlo (10<sup>6</sup> draws, fixed seed; the statistic is scale-free, so τ never enters). Sensitivity: the same with Student-t<sub>5</sub> and t<sub>3</sub> offsets.
- H56 (the model reference is quiet on real views). R against q<sub>M</sub> (Gaussian): S0 false alarms over k = 4–6 ≤ 1.0% on each camera. Refuted above on either.
- H57 (and keeps some power). R against q<sub>M</sub> (Gaussian): ±2% at k = 6, pooled over sign and camera, detected in 15–35% of splits (R against χ²<sub>1</sub>: 53.8%; J against χ²<sub>1</sub>: 17.7%). Refuted outside.
- H58 (χ²<sub>1</sub> is not safe for R until large sets). Under the Gaussian model, P(z<sub>R</sub>² > 6.63) with n<sub>old</sub> = k = n is above 2% for every n ≤ 20. Refuted if it falls to 2% or below at some n ≤ 20.
- Reported, unscored. J against its own q<sub>M</sub>; k = 4 and 5; ±1% and ±4%; the swap; the t<sub>5</sub> and t<sub>3</sub> references; the model's P(z<sub>R</sub>² > 6.63) and q<sub>M</sub> for n<sub>old</sub> = k from 4 to 40; the real false-alarm rate of R against χ²<sub>1</sub> per k next to the model's.
- Caveat set in advance. The model assumes independent per-view offsets, which milestone 4f found only roughly true (the right camera's A13 / DP ratio 1.2–1.7). The splits share 13 views, so no interval is given. H56 and H57 test one rig.
2026-11-13 · Milestone 4g result: on every disjoint split the borrowed threshold catches a 2% focal change in 43%, not 61%; a robust (MAD) jackknife against plain χ²<sub>1</sub> catches it in 54% with nothing borrowed and the same false-alarm rate
- Runs. No new solve: milestone 4f's 8,100 stored 5×5 solves per camera; every 4- to 9-view set had all its delete-one solves converge, so every split is usable. 37,752 splits per camera with an old set of 7 views (25,740 / 10,296 / 1,716 at k = 4 / 5 / 6) and 715 with an old set of 9 and a new set of 4, each scored under 10 scenarios (
experiments/m4g_allsplits/analyse.py, under 5 s). tests/test_m4g.py checked the formulae and that the plain jackknife is quiet (< 5%) with a ratio near 1 on synthetic independent-offset data before the real run. - H52 refuted. At k = 6, pooled over sign and camera, J against χ²<sub>1</sub> detects ±2% in 17.7% (inside 10–28%; 60 splits: 17.5%). But J against the borrowed threshold E<sub>1</sub> detects it in 43.3% (outside 51–71%; 60 splits: 61.3%), and E<sub>1</sub>'s left-camera false alarms over k = 4–6 are 4.4% (not ≥ 5%; 60 splits: 11.1%). Re-estimated on every split, the two cameras' S0 99th percentiles of z² under J are 4.62 (left) and 2.02 (right), against 3.56 and 1.03 from milestone 4d's 180 statistics each.
- Post hoc, computed after H52 failed. This script reproduces milestone 4d's figure exactly on milestone 4b's 60 splits with 4d's stored thresholds (61.25% of 240 decisions). With 4d's thresholds on every split, ±2% is caught in 57.8% and the left camera's false alarms are 10.4% (right 0.5%). So the splits themselves were representative: the drop from 61% to 43% comes from the threshold. A 99th percentile from 180 dependent statistics was too low, and power and false alarms fell together once it was estimated on every split. The borrowed test is also lopsided: ±2% is caught in 56–61% of splits on the left camera and 26–30% on the right.
- H53 refuted. S0, k = 6, real sd of f<sub>new</sub> − f<sub>old</sub> over the root mean predicted variance: R 0.88 (left) and 0.48 (right); J 0.57 and 0.31. The MAD brings the predicted spread from 1.7× to 1.1× the real one on the left camera and from 3.3× to 2.1× on the right, so the right camera's is still twice too wide. At k = 4 and 5 both are much too wide (R 0.25–0.31 at k = 4), because the new set's own delete-one solves sit near the critical set (milestone 4f).
- H54 refuted, narrowly. R against χ²<sub>1</sub>, S0 over k = 4–6: 3.4% (left) and 1.2% (right) false alarms; at k = 6 alone, 3.1% and 0.2%. Pooled over both cameras and k = 4–6 it is 2.3%, the same as the borrowed threshold's 2.4% and well above J's 0.3%.
- H55 held. R against χ²<sub>1</sub>, ±2% at k = 6 pooled: 53.8% (left 50.0–50.1%, right 54.3–61.0%), 10 points above the borrowed threshold and 3× the plain jackknife, with nothing borrowed from another unit.
- Unscored. At k = 6, R catches ±1% in 15.4–16.4% and ±4% in 92.7–94.1% (J: 1.5–2.5% and 69.4–75.4%; E<sub>1</sub>: 13.3–13.4% and 90.7–92.5%). At k = 5 R catches ±2% in 32.0% pooled (E<sub>1</sub> 27.7%, J 8.9%); at k = 4, 11.9–12.0%. An old set of 9 and a new set of 4 is no better than 7 + 4 (R ±2%: 13.8–15.0%): the new set's size limits the test. f alone cannot see the unit swap (0.67% in f): R fires on 1.6–2.6% of swap splits, as on unchanged ones. Harmful adoptions under milestone 4's rule: R 0.28–0.89% of decisions per design, E<sub>1</sub> 0.27–0.75%, J 0.06–0.20%.
- Sensitivity. Every 7 + 6 split uses all 13 views, so leaving a view out leaves no k = 6 split; the pre-registered leave-one-view-out range is undefined there (the gap was found on the first run). At k = 5 the ranges over the 13 left-out views are: R ±2% 23.6–37.3%, E<sub>1</sub> 18.8–35.3%, J 5.1–11.4%. False alarms (k = 4–6): R 2.6–4.0% (left), 0.9–1.7% (right); E<sub>1</sub> left 1.2–7.4%.
- Reading. Two corrections and one gain. First, the paper's headline power for the borrowed-threshold test (61%) was flattered by a noisy threshold from 180 statistics; on every split it is 43%, and the paper has to say so. Second, a robust jackknife is not an honest error bar (the right camera's is still 2× wide), so H53 fails. But against a plain χ²<sub>1</sub> it gives the best test this programme has found: a 2% focal change is caught in 54% of 6-view updates, at the same false-alarm rate as the borrowed threshold, evenly on both cameras, with nothing taken from another unit. For a robot, that removes the need for a sibling unit or a commissioning-time tail estimate at 6 new views. It does not remove the dependence on set size: at 4 new views every variant is weak (13% at most).
- Caveat. One rig, one board, 13 views; the splits are every combination of the same views, so they are far from independent and no interval is given. R's advantage was predicted (H55), but its 3.4% left-camera false alarms are just above the pre-registered bar, and a second unit could put it on either side.
2026-11-13 · Milestone 4g planned: every disjoint split, and a robust jackknife that needs nothing borrowed
Written before any milestone-4g statistic was computed. Milestone 4f solved every subset of 3 to 12 of the 13 views on 5×5 corners and found that the plain view jackknife overstates f's spread mostly because a few delete-one solves of a small set land near the critical set; the median of the per-set jackknife variances was 2–3× closer to the real spread than their mean at 4 views. Two things follow that need no new solve. First, the update test of milestones 4c–4d was scored on 60 splits per k; with every subset stored, every disjoint (old, new) split can be scored, which is the "pre-registered rerun on fresh splits" the paper draft's Section 10 asks for. Second, a per-set robust jackknife (the MAD of the n delete-one focal lengths in place of their standard deviation) may give the f-alone test a usable scale from the unit's own views, where milestone 4d had to borrow a threshold from the other camera and milestone 4e's split-halves were deaf.
- Data. Milestone 4f's stored 5×5 solves only (
experiments/m4f_setsize/results.json.gz, 8,100 per camera); no new solve. experiments/m4g_allsplits/analyse.py. - Splits. Every old set of 7 views and every disjoint new set of k = 4, 5, 6 views from the 13: 25,740, 10,296 and 1,716 splits per camera. Scenarios as milestone 4c: S0 unchanged; S1 f of the new set scaled by s ∈ {0.96, 0.98, 0.99, 0.995, 1.005, 1.01, 1.02, 1.04} (its variance by s²); S3 swap (the old set from the other camera). A set is usable if it and at least n − 1 of its n delete-one solves converged.
- Variances of f per set (from its delete-one focal lengths f<sub>i</sub>, m usable): J = (m − 1)/m Σ(f<sub>i</sub> − f̄)² (milestones 4b–4f); R = (m − 1)²/m · (1.4826 · median|f<sub>i</sub> − median f<sub>i</sub>|)², the same with the MAD in place of the standard deviation. Statistic z² = (f<sub>new</sub> − f<sub>old</sub>)² / (v<sub>old</sub> + s²v<sub>new</sub>), against χ²<sub>1</sub> at 1% (6.63). A zero denominator counts as fired. E<sub>1</sub>: milestone 4c's borrowed threshold (the other camera's S0 99th percentile of z² under J, k ≥ 4), recomputed on all splits.
- H52 (the 60-split figures survive every split). Pooled over cameras at k = 6: J against χ²<sub>1</sub> detects ±2% in 10–28% of splits (60 splits: 17.5%); J against E<sub>1</sub> detects ±2% in 51–71% (61.3%); and E<sub>1</sub>'s S0 false alarms on the left camera (k ≥ 4) are ≥ 5% (11.1%). Refuted if any fails.
- H53 (R has the right scale). S0, k = 6: the real sd of (f<sub>new</sub> − f<sub>old</sub>) about each camera's mean, over √(mean(v<sub>old</sub> + v<sub>new</sub>)) with R, lies in 0.67–1.5 on both cameras. Refuted outside on either. (Under J the same ratio is predicted near 0.4–0.7.)
- H54 (R is quiet). R against χ²<sub>1</sub>: S0 false alarms over k = 4–6 ≤ 3% on each camera. Refuted above on either.
- H55 (R has power with nothing borrowed). R against χ²<sub>1</sub>: ±2% at k = 6, pooled over sign and camera, detected in ≥ 40% of splits. Refuted below.
- Reported, unscored. k = 4 and 5; ±0.5, ±1 and ±4%; the swap; an old set of 9 views with a new set of 4 (715 splits); harmful adoptions under milestone 4's rule; the rates with every split that uses view v left out, for each of the 13 v (min–max).
- Caveat set in advance. The splits are every combination of the same 13 views, so they are far from independent: rates are reported without intervals, and the leave-one-view-out range is the only spread given. A MAD of 6 or 7 numbers is itself noisy, so R's z² may have a heavier tail than χ²<sub>1</sub> even if its scale is right (H53 can hold while H54 fails).
2026-11-13 · Milestone 4f result: the jackknife's overstatement of f's spread falls from 7–8× at 4 views to 1.1–1.4× at 10, and the solver's σ is 2–6× small at every size
- Runs. 10,854 new solves on 5×5 corners, merged with milestones 4d and 4e's 5,346: every subset of 3 to 12 of the 13 views on both cameras (16,172 view sets, plus the two all-view solves). All converged except 2 of the right camera's 286 three-view sets.
analyse.py was checked on synthetic independent-offset data first (tests/test_m4f.py: every ratio reads 0.85–1.2 there), then run once on the real solves. - H48 held. Jackknife over real (disjoint pairs): 7.8 at 4 views and 1.8 at 6 (left), 8.2 and 3.1 (right). Against A13 the ratio falls at every step from 4 to 10 views on both cameras (Spearman −1.0): left 6.7, 3.0, 1.6, 1.4, 1.3, 1.2, 1.1; right 6.7, 4.6, 1.9, 1.8, 1.7, 1.6, 1.4.
- H49 refuted. Left: 1.28, 1.17, 1.10 at 8, 9 and 10 views (inside 0.67–1.5). Right: 1.68, 1.57, 1.43 (outside at 8 and 9). The right camera needs about 10 views before the jackknife comes within 1.5×.
- H50 refuted. A13 over disjoint pairs: 1.17, 1.16, 1.17 at 4, 5, 6 views (left, inside 0.8–1.25); 1.22, 1.40, 1.66 (right, outside at 5 and 6). On robust scales the gap shrinks (0.94–1.01 left, 1.13–1.29 right), so heavy tails explain part of the right camera's gap but not all of it. On the right camera the views do not add independent offsets, and A13 overstates the real spread there. The jackknife's real overstatement on that camera above 6 views is therefore probably larger than the A13 ratios say.
- H51 held. The solver's own σ over real (A13), n = 4..10: 0.17–0.24 (left), 0.19–0.59 (right). On unbiased corners it is 4–6× too small on the left camera, worse than the 2× milestone 1 found on 11×11 corners. That is consistent with milestone 6 (better corners make σ less honest).
- Unscored. At 11 and 12 views the jackknife ratio is 1.04 / 1.01 (left) and 1.28 / 1.16 (right), but A13 divides by 1 − n/13 ≤ 0.15 there. Robust jackknife ratios (median v<sub>J</sub>) are much closer to 1 at small n: 2.4 (left) and 4.1 (right) at 4 views against 6.7 for the mean, so a few delete-one solves landing near the critical set make most of the small-set overstatement, as milestone 4e guessed. Leave-one-view-out ranges of the jackknife ratio (A13): 4.9–9.5 / 4.5–12.8 at 4 views; 0.86–1.39 / 1.46–1.92 at 8; 0.42–1.32 / 0.99–1.71 at 10 (left / right). Real per-set sd (A13) falls from 11–14% of f at 3 views to 2.1–2.7% at 4, 0.6–0.7% at 6 and 0.24–0.40% at 10.
- Reading. Milestone 4e's guess holds: the jackknife's overstatement is mostly a small-set effect from near-critical delete-one solves, and it shrinks steadily as the set grows. It is within 1.5× from about 7 views on the left camera and 10 on the right. For a deployed robot this means two things. Below 6 views neither σ is usable: the solver's is 4–6× too small and the jackknife's 2–8× too large. With 10 or more logged views, the plain jackknife is a usable, mildly conservative error bar without a borrowed threshold. Milestone 4e's split-half scale would also become usable then (5 + 5 halves), and the two can be checked against each other on one unit. The second camera shows that "independent per-view offsets" is only an approximation, so a submission should report the disjoint-pair reference wherever 2n ≤ the views available.
- Caveat. One stereo rig, one board, 13 views shared by every set; the per-n figures are far from independent and the leave-one-view-out ranges are the only uncertainty given. 5×5 corners only.
2026-11-13 · Milestone 4f planned: does the jackknife stop overstating f's spread once the view set leaves the critical region?
Written while the milestone-4f solves ran, before any of them was analysed. Milestone 4e found, post hoc, that f's spread per view falls 2–5× between 3 and 5 views and then roughly follows 1/n. It guessed that the same effect is why the leave-one-view-out jackknife overstated the real set-to-set spread of 7-view sets (1.8× and 2.3× on 5×5 corners): deleting one view from a small set lands near the critical set, so a few delete-one solves swing far and inflate the variance. If the guess is right, the overstatement should shrink as the old set grows, and a robot that logs about 10 views could use the jackknife as it is. Here every subset of 3 to 12 of the 13 views is solved on each camera (8,086 per camera, experiments/m4f_setsize/run.py; milestones 4d and 4e's 5×5 solves reused), on 5×5 corners only (milestone 6b's unbiased localiser), with milestone 4b's solver, reference views and start.
- Quantities (per camera, relative to the all-view f<sub>13</sub>, converged solves only): - real sd, disjoint pairs (DP), n = 3..6: √(mean d²/2) over every unordered pair of disjoint n-view sets, d = (f<sub>A</sub> − f<sub>B</sub>)/f<sub>13</sub>. Model-free; needs 2n ≤ 13. - real sd, around all views (A13), any n: √(mean (f<sub>S</sub> − f<sub>13</sub>)² / (1 − n/13)) over every n-view set S. Exact if each view adds an independent offset (then var(f<sub>S</sub> − f<sub>13</sub>) = var(f<sub>S</sub>)(1 − n/13)); the only real reference above 6 views. - jackknife J(n) = √(mean v<sub>J</sub>) over n-view sets with all n delete-one solves converged, v<sub>J</sub> = (n−1)/n Σ(f<sub>i</sub> − f̄)²/f<sub>13</sub>² (milestone 4b's
jackknife_cov, f entry). - reported R(n) = √(mean σ<sub>f</sub>²)/f<sub>13</sub> (the solver's own covariance). - ratio ρ<sub>J</sub>(n) = J(n)/real(n); reference DP for n ≤ 6 where stated, A13 otherwise. - H48 (overstatement falls with set size). On each camera ρ<sub>J</sub>(4) > ρ<sub>J</sub>(6) against DP, and the Spearman correlation of ρ<sub>J</sub>(n) (A13) with n over n = 4..10 is ≤ −0.8. Refuted if either fails on either camera.
- H49 (honest from about 8 views). ρ<sub>J</sub>(n) (A13) lies in 0.67–1.5 for each of n = 8, 9, 10 on both cameras. Refuted if any falls outside.
- H50 (the independent-offset model holds where it can be checked). At n = 4, 5, 6, A13 over DP lies in 0.8–1.25 on both cameras. Refuted otherwise (and then H48's and H49's A13 reference is suspect above 6 views).
- H51 (the solver's σ stays small at every size). R(n)/real(n) (A13) ≤ 0.67 for every n = 4..10 on both cameras (milestone 1: about 2× small on 13 views). Refuted if any exceeds.
- Reported, unscored. n = 3, 11, 12; robust versions (1.4826 MAD for the real spread, median v<sub>J</sub> for the jackknife); convergence counts per n; a leave-one-view-out sensitivity (each ratio recomputed with every set holding view v dropped, for each of the 13 v; min–max reported).
- Caveat set in advance. All sets are drawn from the same 13 views and overlap heavily, so the per-n figures are far from independent and carry no honest interval beyond the leave-one-view-out range. The A13 reference at n ≥ 11 divides by 1 − n/13 ≤ 0.15 and is too noisy to score. 5×5 corners only: the other localisers carry a bias (milestone 6b) that adds a shared offset to every set.
2026-11-13 · Milestone 4e result: a threshold from the old set's own 4 + 3 split-halves is quiet but deaf, because f's spread does not shrink as 1/n below 5 views
- Runs. 3,168 new solves (every 3- and 4-view subset of the 13 views not already stored, × 3 localisers × 2 cameras), all with the milestone-4b solver; merged with milestones 4b and 4d, every one of the 171 old sets per camera and localiser had all 35 partitions usable. 14,040 decisions scored with milestone 4c's
score(), unchanged, none left undecided. - H43 held. z<sub>H</sub>² against F<sub>1,6</sub>, 5×5 corners, S0, k ≥ 4: 0.6% (left, 95% 0.1–3.1%) and 0.0% (right) false alarms.
- H44 refuted. ±2% at k = 6 pooled: 0.8% (0.2–3.0%), against 61% for the borrowed threshold. Against χ²<sub>1</sub> it is still only 4.2%; ±4% is caught in 7–9% (F) and 27–31% (χ²<sub>1</sub>).
- H45 refuted. Real sd of (f<sub>new</sub> − f<sub>old</sub>)/f<sub>old</sub> at k = 6 over predicted: 0.67% / 1.94% = 0.34 (left), 0.50% / 2.62% = 0.19 (right). The split-half scale overstates the spread 2.9× and 5.2×, more than the jackknife did (1.8× / 2.3×).
- H46 refuted. S0 99th percentiles of z<sub>H</sub>²: 5.34 (left) against 1.24 (right), a factor 4.3.
- H47 held. Harmful adoptions 0.02% of 4,680 decisions, because the test almost never fires.
- Unscored. Median τ̂ per view (5×5): 2.98% left, 3.96% right (11×11 3.23 / 6.48%, SB 4.30 / 5.11%). On 11×11 and SB corners the picture is the same: scale ratio 0.46 / 0.22 and 0.33 / 0.25, ±2% detected in 0–2%. The unit swap is invisible to f alone (0.0%); the union with W against χ²<sub>3</sub> is W alone (79.4% swap, 1.9% false alarms).
- Why, post hoc (computed after H45 failed). The robust sd (1.4826 MAD) of each subset's f around the all-view f, divided by √(1/n − 1/13) (the part of the gap a subset's own views make if each view adds an independent offset), should be the same at every n. On 5×5 corners it is 3.29% / 5.21% per view at 3 views, 2.15% / 2.17% at 4, 1.68% / 1.35% at 5, 1.67% / 1.04% at 6 and 1.55% / 0.95% at 7 (left / right); the update spread at k = 6 implies 1.20% / 0.91%. Below 5 views the halves sit near the critical set, where milestone 2 found the error grows fastest, so a 3-view half carries 2–5× the per-view spread a 6- or 7-view set does. 1/n holds roughly from 5 views up.
- Reading. Two of five predictions held, and both only because the test is silent. A per-unit scale from the old calibration's own views is the right idea for the threshold not to travel between units, but it needs halves in the regime where 1/n holds: at least 5 views each, so about 10 views in the old set, or halves of an old set enlarged by views a robot logs during normal operation. With 7 views, the borrowed or pooled threshold remains the only usable f-alone test. It may also be part of why milestone 3 found the jackknife very conservative at 4 views (deleting one view from a small set lands near the critical set); that was not tested here.
- Caveat. The 35 partitions of 7 views are dependent, and the 13 views are shared by every split, so the intervals are too narrow; the post hoc per-view figures use subsets that overlap one another heavily.
2026-11-13 · Milestone 4e planned: a threshold from the old calibration's own views, with no sibling unit
Written before any milestone-4e solve was run. Milestone 4d's f-alone update test caught a 2% focal change in 61% of splits on unbiased 5×5 corners, but its threshold was borrowed from the other camera, and it no longer travelled: 11.1% false alarms on the left camera, because the jackknife overstates the real set-to-set spread by different amounts on the two cameras (1.5× and 2.8×). The paper draft names a per-unit tail estimate as what a submission needs. Here the scale comes from the old set itself, which a robot always has: each old set of 7 views is split into 4 + 3 views in all 35 ways, and the spread of the two halves' focal lengths, scaled by 1/n, predicts the spread between the old set and a new set of k views. Nothing is borrowed from another unit and the new views are never used to set the threshold.
- Solves. Every 3- and 4-view subset of the 13 views (1,001 per camera) that milestones 4b and 4d did not already solve, on all three localisers (
experiments/m4e_selfscale/run.py: the same solver, k<sub>1</sub> model, reference views and start). - Statistic. For an old set O and each partition (A, B) with both halves converged, r = (f<sub>A</sub> − f<sub>B</sub>)/f<sub>O</sub>. The per-view scale is τ̂² = median(r²) / 0.45494 / (1/4 + 1/3) (the median of χ²<sub>1</sub> makes it robust to a near-critical half). Predicted variance of the relative update v = τ̂²(1/7 + 1/k), and z<sub>H</sub>² = ((f<sub>new</sub> − f<sub>old</sub>)/f<sub>old</sub>)² / v. A decision needs at least 20 of the 35 partitions converged. Splits, disturbances (S1 focal change, S0 unchanged, S2 principal-point shift, S3 unit swap, where the old set and so its scale come from the other camera), k ≥ 4 scoring and the adoption rule are milestone 4c's.
- Reference. Primary: F<sub>1,6</sub> at 1% (13.75), since τ̂ rests on 7 views. Secondary, also scored: χ²<sub>1</sub> (6.63).
- H43 (quiet with nothing borrowed). On 5×5 corners, z<sub>H</sub>² against F<sub>1,6</sub>: S0 false alarms (k ≥ 4) ≤ 3% on each camera. Refuted if above on either (the borrowed threshold: 11.1% left).
- H44 (power kept). On 5×5 corners, F<sub>1,6</sub>: a ±2% focal change at k = 6, pooled over sign and camera, detected in ≥ 40% of splits (borrowed threshold 61%). Refuted below.
- H45 (the scale is right). On 5×5 corners, S0, k = 6: the real standard deviation of (f<sub>new</sub> − f<sub>old</sub>)/f<sub>old</sub> about each camera's mean, divided by the root mean predicted v, lies within 0.67–1.5 on both cameras (the jackknife: 0.67/1.23 = 0.54 and 0.51/1.17 = 0.44). Refuted outside on either.
- H46 (the tail travels). On 5×5 corners, the two cameras' S0 99th percentiles of z<sub>H</sub>² (k ≥ 4) are within a factor of 2 of each other (jackknife z<sub>f</sub>²: 3.56 against 1.03, 3.5×). Refuted otherwise.
- H47 (no harm). On 5×5 corners, F<sub>1,6</sub>: harmful adoptions over all scenarios (k ≥ 4) ≤ 1.5% (borrowed threshold 1.3%). Refuted above.
- Reported, unscored. The same on 11×11 and SB corners; χ²<sub>1</sub>; ±1% and ±4%; the principal-point shift and the unit swap; a union with milestone 4b's W against χ²<sub>3</sub> for the principal point; comparisons with the borrowed (E<sub>1</sub>) and oracle thresholds.
- Caveat set in advance. The 1/n scaling assumes each view adds an independent offset; 3- and 4-view halves sit nearer the critical set than 7-view sets, which would inflate τ̂ and make the test conservative. The 35 partitions of 7 views are not independent, and the splits reuse 13 views, so the intervals are too narrow.
2026-11-13 · Milestone 7: first full paper draft, every number tied to a stored result
- What.
paper/draft.md, an RA-L-shaped draft (abstract, ten sections, a results table) that pulls milestones 1–6b into one argument: target-free BA matches Zhang to 0.15–0.24%, its reported σ is 2× small, a view jackknife makes the update test quiet (1.7–2.5% false alarms), the half-width end squares bias two of three corner localisers, and on unbiased 5×5 corners the f-alone test catches a 2% focal change in 61% of splits. paper/references.bib holds the 16 works cited, all marked unchecked (the DOI and arXiv hosts refused this session's proxy). - Checked.
paper/check_numbers.py (stdlib) ties 82 quoted figures to the experiments' summary.json files (56) or, where a figure was printed but not stored, to this log's result entries (26); each check also requires the figure to appear in the draft. 82 of 82 pass. - Corrections found while drafting. None to the results. One wording slip in the draft's own first pass (the principal-point shifts are 2, 5 and 10 px, not "5–20 px") was caught against
m4b_viewcov/analyse.py before commit. - What the draft says a submission still needs (its Section 10): a second camera and board with a physically changed lens and enough views per unit to estimate each unit's own tail on views the test never touches (milestone 6, owner step); natural-feature correspondences (milestone 3b); a pre-registered rerun of the 5×5 test on fresh splits with a pooled threshold; checked citations; cutting to 8 pages.
- Looked at and set aside. OpenCV's
opencv_extra test data was checked for a larger pinhole calibration set to test a per-unit threshold: its biggest stereo set (fisheye/calib-3_stereo_from_JY, 34 pairs) is a fisheye rig the k<sub>1</sub> pinhole model does not fit, and the rest are 6–15 views. Not used.
2026-11-13 · Milestone 4d result: unbiased 5×5 corners nearly double the f-alone test's power for a 2% focal change (61% against 34–35%), but the sibling's threshold stops travelling (11% false alarms on the left camera)
- Runs. 4,290 solves on 5×5-window corners (milestone 4b's 2,145 view sets per camera), 1 not converged (a 3-view set); re-solving the all-view set on 11×11 and SB corners reproduced milestone 4b's stored f, c<sub>x</sub>, c<sub>y</sub> and k1 exactly (difference 0.0 on all four fits, OpenCV 5.0.0). Scored with milestone 4c's
score() (factored out of its main, output byte-identical), 4,680 decisions on 5×5 corners, k ≥ 4. - H38 held. W against χ²<sub>3</sub> with J: 1.7% false alarms pooled (95% 0.8–3.6%), 2.2% left, 1.1% right.
- H39 held. Standard deviation of (f<sub>new</sub> − f<sub>old</sub>)/f on unchanged splits at k = 6: 0.59% on 5×5 corners, against 1.02% (11×11) and 0.86% (SB). Per camera 0.67% / 0.51% (11×11 1.03 / 1.01%, SB 0.91 / 0.81%).
- H40 held. z<sub>f</sub>² against the other camera's 99th percentile detects ±2% at k = 6 in 61.3% of splits (55–67%; −2% 64%, +2% 58%), against 34.2% (11×11) and 35.4% (SB).
- H41 refuted. The borrowed threshold over-alarms on the left camera: 11.1% (7.3–16.5%), right 0.0%. The left camera's own 99th percentile of z<sub>f</sub>² is 3.56, the right's 1.03: the jackknife σ shrank less than the real spread, and by different amounts (the jackknife σ of the difference is 1.23% / 1.17% of f against a real spread of 0.67% / 0.51%, rms z 0.65 / 0.36).
- H42 held. The real unit swap, W against χ²<sub>3</sub> with J, k ≥ 4: 79.4% (75–83%), against 30.3% (11×11) and 74.7% (SB).
- Unscored. At k = 6 on 5×5 corners: f alone against the borrowed threshold detects ±1% in 28% / 18% (11×11: 4 / 10%; SB: 12 / 7%) and ±4% in 98% / 97%; the oracle threshold gets ±2% in 66% / 62%, so the borrowed one loses little power, only calibration. W against χ²<sub>3</sub> detects ±2% in 20% / 25% (11×11 8 / 9%, SB 12 / 12%) and a 10 px principal-point shift in 27% (6% / 12%). Harmful adoptions over all scenarios: 1.3% (f alone, borrowed threshold), 1.5% (W, χ²), 2.3% (the exploratory union, whose left-camera false alarms are 12.8%). The all-view target-free f on 5×5 corners is 534.3 / 535.8 px, +0.27% / −0.08% from Zhang on the same corners (532.9 / 536.2).
- Reading. Four of five predictions held. The half-width end squares' localiser bias was a large part of the set-to-set offset that made the test deaf: with corners that bias does not reach, the focal-length spread between disjoint view sets falls by a third to two fifths, and the f-alone test catches a 2% focal change three times in five and a 1% change about one time in four, where before it caught 2% one time in three and 1% hardly at all. The cost is calibration of the threshold: the jackknife now overstates the spread 1.5× on one camera and 2.8× on the other, so one sibling's tail is no longer the other's. A deployable version needs a threshold pooled over several sibling units or a per-unit tail estimate, not one borrowed unit.
- Caveat. The 99th percentiles still rest on about two of 180 statistics from splits that reuse 13 views, so the left camera's 11% might shrink with more splits; the two cameras share every board pose. The H41 explanation (unequal jackknife shrinkage) was computed after the hypothesis failed.
2026-11-13 · Milestone 4d planned: does the unbiased 5×5 window make the update test sharper?
Written before any milestone-4d solve was run. Milestone 6b found that this board's half-width end squares pull the 11×11 cornerSubPix window's outer-ring corners by up to 8 px and bias Zhang's f by +0.6–0.9%, and SB's by −0.1 to −0.4%, while a 5×5 window is unbiased (f within 0.04% on renders). Milestones 4b–4c ran the update test only on 11×11 and SB corners. The bias depends on each view's pose, so it should differ from one view set to the next and add to the set-to-set offset that made the test deaf to a 2% focal change. Here milestone 4b's solves are repeated on 5×5 corners (experiments/m4d_window/run.py: the same 2,145 view sets per camera, solver, reference views and start) and scored with milestone 4c's statistics and references, unchanged.
- H38 (the jackknife stays honest). W against χ²<sub>3</sub> with the jackknife covariance J: S0 false alarms (k ≥ 4) ≤ 3% pooled and ≤ 5% on each camera. Refuted otherwise.
- H39 (less set-to-set offset). The standard deviation of (f<sub>new</sub> − f<sub>old</sub>)/f on unchanged splits at k = 6, pooled over cameras after removing each camera's mean, is smaller on 5×5 corners than on both 11×11 and SB corners. Refuted if it is not smaller than both.
- H40 (more power for focal change). z<sub>f</sub>² against the other camera's S0 99th percentile (E<sub>1</sub>) detects a ±2% focal change at k = 6, pooled over sign and camera, in ≥ 40% of splits (11×11: 34%, SB: 35%). Refuted below 40%.
- H41 (the borrowed threshold travels). E<sub>1</sub> S0 false alarms (k ≥ 4) ≤ 5% on each camera. Refuted if above on either (SB's left camera was at 5.0%).
- H42 (the swap). W against χ²<sub>3</sub> with J detects the real unit swap S3 (k ≥ 4, both directions) in ≥ 50% of splits (11×11: 30%, SB: 75%). Refuted below.
- Reported, unscored. The full power table and harmful adoptions for every reference; thresholds; the all-view target-free f on 5×5 corners against 5×5 Zhang (532.9 / 536.2 px, milestone 6b's best estimate); a check that this environment reproduces milestone 4b's stored all-view solutions on 11×11 and SB corners (OpenCV here is 5.0.0).
- Caveat carried over. The splits reuse 13 views, so the Wilson intervals are too narrow, and a 99th percentile from 180 statistics rests on about two of them.
2026-11-13 · Milestone 6b result: the printed board's end squares are half width, which biases the 11×11 window's f by +0.6–0.9% and SB's by −0.1 to −0.4%; a 5×5 window is right
- Runs. 26 renders at the real poses (truth: each camera's all-view 11×11 Zhang fit with k1 + k2), 5 settings × 20 seeds × 2 cameras, 3 localisers, 600 flat-board Zhang and 600 target-free fits; all detections and fits succeeded except one SB render. Matched setting, calibrated so the renders' own statistics equal the real images': render blur σ 0.80 / 0.81 px and pre-JPEG noise 1.31 / 1.50 grey levels (left / right) give edge σ 0.949 / 0.960 px (real 0.948 / 0.957) and noise 0.54 / 0.63 (real 0.52 / 0.61). Tests check the renderer puts corners at the projected points (< 0.1 px) and the statistics recover known blur and noise.
- H33 refuted. On the full-square renders no outer-ring corner differs by more than 1 px between 11×11 and SB (max 0.19 / 0.36 px), against 22 / 30 on the real images (max 7.7 / 5.6 px).
- H34 refuted. All three localisers are within 0.035–0.060 px rms of the truth on the ideal pattern; 11×11 is lowest on the ring of both cameras. The error is almost all systematic (the per-corner mean over 20 seeds has the same rms), not noise.
- H35 refuted, second clause. SB's f is within 0.025% of the truth (the 0.2% clause held), but so is every localiser's (11×11: −0.004% / +0.030%): on an ideal pattern none of them moves f, so the real 0.8–1.3% shift is not localiser behaviour on a clean board.
- H36 refuted. Target-free BA on SB corners: rms z of f against the truth 1.55 / 1.80. The reported σ is 2–2.5× larger than the seed-to-seed spread (spread/σ 0.46 / 0.39), but a fixed bias of −0.02% / +0.04% of f, the localiser's systematic error, is 1.5–2σ. A sandwich clustered by view absorbs it (rms z 0.75 / 0.98).
- H37 held. Adjacent along-column residual correlation −0.09 to +0.08 for every localiser on the renders (real 0.32–0.56). Unscored: SB's along-row correlation on the right camera is 0.26–0.35 on renders too.
- Unscored, the main finding (written after H33 was refuted). Magnified crops of the six worst real corners show the 11×11 corner 4–8 px off the visible corner, towards the board's edge, with SB and 5×5 on it. Every real corner with an 11×11–SB gap over 1 px is in grid column 0 or 8. Measured outward along the grid lines (
outer.py), the real board's end squares in those two columns reach only 0.49–0.53 of the corner spacing past the last corner (10–90%: 0.35–0.59), against 0.94–0.97 on the other two sides; on the renders the same measurement reads 0.95–1.01. Re-rendered with half-width end squares (ablate.py, 5 seeds), the 11×11 window misplaces 12–18 ring corners per 13-view set by more than 1 px (max 3.5–9.6 px) and biases Zhang's f by +0.43% to +1.47% depending on the white margin and the surround (real 11×11 minus 5×5: +0.64% / +0.91%). SB is biased the other way, −0.32% to −0.43% (real SB minus 5×5: −0.12% / −0.39%), and fails to detect the board when the surround is narrow and dark. The 5×5 window is untouched by any surround (f within 0.04% of the truth, ring rms 0.05–0.06 px). Full-width end squares with any margin or surround leave all three localisers unbiased, so the half-width squares are the cause. - Reading. One of five predictions held. The question milestone 6 left, which localiser is right, has an answer for this board: the 5×5 window. Zhang's f on 5×5 corners, 532.9 px (left) and 536.2 px (right), is the best estimate; the 11×11 values used in milestones 1–5 (536.3 / 541.1) are about 0.6–0.9% high, and SB's (532.2 / 534.1), used in milestones 4b–4c, 0.1–0.4% low. The general lesson for calibration practice: a corner whose sub-pixel window reaches past a cut-short end square is pulled towards the board's edge, so either print full end squares and a margin of at least one window, or size the window below the shortest end square.
- Caveat. The renders leave out the bend, the clipboard and background clutter, vignetting and the camera's own noise; the half-width ablation fixes the reach at 0.5 for every row, where the real reach varies (0.35–0.59). The ablation and the reach measurement were added after H33 failed and were not pre-registered.
2026-11-13 · Milestone 6b planned: which corner localiser is right? Renders with a known camera at the real poses
Written before any image was rendered. Milestone 6 found that the corner localiser alone moves flat-board Zhang's f by 0.9–1.3% (4.8 / 7.2 px) and target-free f by 1.8 / 4.2 px, that cornerSubPix 11×11 and findChessboardCornersSB disagree by more than 1 px on 52 corners, all on the grid's outer ring, and that no ground truth said which is right. Here the truth is known: each camera's 13 views are re-rendered from a known camera and known poses, and every localiser is scored against the exact projections.
- Renderer. The true camera of each side is its all-view flat-board Zhang fit on 11×11 corners with k1 and k2 (OpenCV model, no tangential terms), and the 13 board poses are that fit's poses, so each render puts the board where it was in the real image. Each 640×480 pixel is the mean of a 4×4 grid of sub-pixel rays, each undistorted (fixed-point inversion of the radial model), intersected with the board plane and shaded: a flat 10×7-square checkerboard (9×6 inner corners), a white margin of 0.5 square around it (the real board's white border, read off the images), a uniform mid-grey background beyond. Then a Gaussian blur, additive Gaussian noise, 8-bit rounding and JPEG encoding.
- Matched setting, fixed before any localiser is scored on a render. Black and white levels are the medians of the real images' square interiors near detected corners; blur σ is fitted to the real images' edge profiles (an erf fit across the board's inner edges); noise σ is the rms of a 3×3 Laplacian in flat square interiors, rescaled to a pixel σ; JPEG quality 90. Also swept: blur 0.5× and 2× the matched value, and noise 0 and 2×. 20 noise seeds per setting.
- Localisers. The three of milestone 6:
cornerSubPix 11×11 and 5×5 half-windows, and SB with CALIB_CB_ACCURACY. Error = detected minus true projection, matched by corner index (truth ordering). - H33 (the render reproduces the ring). In the matched setting, at least one outer-ring corner per camera differs by more than 1 px between 11×11 and SB, as on the real images. Refuted if none does on either camera: then the real ring disagreement comes from something the render leaves out (the clipboard edge, clutter, lighting), not from the ideal pattern.
- H34 (SB is closest to the truth). Matched setting, pooled over seeds and views: SB's rms corner error is below 11×11's, on the outer ring and in the interior, on both cameras. Refuted if 11×11 or 5×5 is lower on either subset on either camera.
- H35 (the localiser picks the focal length). Flat-board Zhang (k1 + k2, as the truth) on SB corners recovers f within 0.2% of the truth on both cameras (median over seeds), and 11×11's |f error| is larger than SB's on both. Refuted if SB is off by more than 0.2% on either camera, or 11×11 is no worse on either.
- H36 (σ is honest when the world is ideal). Target-free all-view BA (k1 + k2) on SB corners, z = (f − f<sub>true</sub>)/σ<sub>f</sub> over the 20 seeds: rms z ≤ 1.5 on both cameras. Refuted if above 1.5 on either. On the real images the nested rms z was 3.12 with SB corners (milestone 6); if renders are honest the 2–3× comes from the world, not the localiser on an ideal pattern.
- H37 (no column structure on an ideal pattern). In the matched setting, the target-free all-view residuals' adjacent along-column correlation is below 0.2 for every localiser on both cameras (real: 0.44–0.56 with 11×11, 0.32 / 0.56 with SB). Refuted if 0.2 or more for any localiser on either camera.
- Reported, unscored. All five predictions in the blur and noise sweep; per-ring error maps; target-free f error per localiser; whether the real-image f shift between localisers (milestone 6) is reproduced.
- Caveat set in advance. A render is a model of the image, not the image: it leaves out the clipboard's rounded edge, background clutter, vignetting, defocus that varies with depth, the bend milestones 5–6 found, and the sensor's real noise and demosaicing. A localiser that wins on renders is right for the ideal pattern; a real-image gap the renders do not reproduce points at what they leave out.
2026-11-12 · Milestone 4c result: the 3-parameter test is information-limited, but testing f alone against a borrowed tail threshold catches a 2% focal change 3–4× as often
- Runs. No new solve: milestone 4b's 8,580 stored solves, its splits and disturbances, k ≥ 4, 9,360 decisions (360 unchanged splits per detector). Satterthwaite ν on unchanged splits: median 6.2 (11×11) / 5.9 (SB), 10–90% 3.0–10.1, so the Hotelling threshold has median 68 / 79 against χ²<sub>3</sub>'s 11.3. Empirical 99th percentiles of W on unchanged splits: 17.5 / 14.7 (11×11, left / right), 9.8 / 15.2 (SB); of z<sub>f</sub>²: 2.03 / 1.69 and 4.22 / 1.95, against χ²<sub>1</sub>'s 6.63.
- H28 held. F reference: 0 false alarms in 360 on both detectors (95% 0–1.1%).
- H29 held. F keeps 19% (11×11) and 22% (SB) of χ²'s ±4% detections at k = 6 (8% against 43%; 13% against 56%).
- H30 held. The other camera's W threshold: 2.2% / 0.6% false alarms (11×11, left / right), 0.6% / 2.8% (SB).
- H31 held. The oracle threshold on W detects ±2% at k = 6 in 4.6% (11×11) and 10.0% (SB) of splits: no threshold on the 3-parameter statistic rescues it.
- H32 refuted. z<sub>f</sub>² against χ²<sub>1</sub> detects ±2% in 3.3% (11×11) and 7.9% (SB), fewer than W against χ²<sub>3</sub> (8.3% / 11.7%); its false alarms are 0.0% / 0.3%. Against χ²<sub>1</sub> the f-only test is too quiet, because the jackknife overstates f's variance by about 3× (99th percentile 1.7–4.2 where χ²<sub>1</sub> says 6.6).
- Unscored, and the main finding. z<sub>f</sub>² against the other camera's 99th percentile (E<sub>1</sub>) detects ±2% at k = 6 in 41% / 28% (11×11, −2% / +2%) and 40% / 31% (SB), and ±4% in 78–85%, with false alarms 2.2% / 0.6% (11×11) and 5.0% / 0.0% (SB, left / right). The same camera's own threshold (oracle) gives 43% / 28% and 38% / 33%: the borrowed threshold loses almost nothing. A 1% change stays invisible (4–12%). The f-only test cannot see the unit swap (1–4%), which lives in c<sub>x</sub>; an exploratory union (W or z<sub>f</sub>², both thresholds borrowed, not pre-registered) keeps the swap (21% / 73% at k ≥ 4) and the ±2% power (29–43% at k = 6), at 2.8% / 4.2% false alarms (left camera 4.4% / 5.6%).
- Harmful adoptions (milestone 4's rule, all scenarios, k ≥ 4): f-only with E<sub>1</sub> 0.47% (11×11) and 0.96% (SB); the union 0.68% / 2.29%.
- Reading. Four of five predictions held. The jackknife test's deafness had two sources: for the 3-parameter statistic it is information (the oracle buys nothing), but for f the χ² reference is badly mis-scaled, and a tail threshold borrowed from a sibling unit recovers most of the oracle's power. The practical test is two statistics, f against a borrowed threshold for focal change and W for principal-point moves.
- Caveat. The thresholds are 99th percentiles of 180 statistics from splits that reuse 13 views; the E<sub>1</sub> false-alarm rate on the SB left camera (5.0%) shows the borrowed tail does not always match. The union and the E<sub>1</sub> power were not pre-registered.
2026-11-12 · Milestone 4c planned: is the jackknife test deaf because of its χ² reference or because 4–6 views carry too little information?
Written before any milestone-4c number was computed. Milestone 4b's jackknife test (J) was quiet (1.7–2.5% false alarms at a nominal 1%) but caught a ±2% focal change in only 7–12% of splits with 6 new views. Two explanations predict different fixes: the χ²<sub>3</sub> reference is wrong for a covariance estimated from 4–7 delete-one solves (then a better reference or a narrower test buys power back), or the views simply do not carry the information (then no threshold helps). Milestone 4c answers it from milestone 4b's 8,580 stored solves (m4b_viewcov/results.json.gz), with no new solve, on the same splits, disturbances, detectors and k ≥ 4 scoring.
- Statistic. W = dᵀ(J<sub>old</sub> + s²J<sub>new</sub>)⁻¹d on d = θ<sub>new</sub> − θ<sub>old</sub>, θ = (f, c<sub>x</sub>, c<sub>y</sub>), as in 4b; and the one-parameter z<sub>f</sub>² on f alone.
- Four references for W at 1%. (χ²) 11.34, as 4b. (F) a Hotelling-type reference with Satterthwaite degrees of freedom from the f variances, ν = (a + b)² / (a²/(m<sub>old</sub> − 1) + b²/(m<sub>new</sub> − 1)), a = J<sub>old,ff</sub>, b = s²J<sub>new,ff</sub>, m the delete-one solves that converged; threshold 3ν/(ν − 2) · F<sub>3, ν−2</sub>(0.99). (E) the other camera's empirical 99th percentile of W on unchanged splits (S0), same detector, k ≥ 4 pooled. (O, oracle) the same camera's own S0 99th percentile, k ≥ 4 pooled: false alarms ≈ 1% by construction, so it is an upper bound on the power any recalibration of the threshold can buy, not a usable test.
- Two references for z<sub>f</sub>². χ²<sub>1</sub> (6.63) and F<sub>1, ν</sub>(0.99).
- H28 (F is safe). J with the F reference: S0 false alarms (k ≥ 4, pooled) ≤ 1% on both detectors. Refuted if above 1% on either.
- H29 (F is too strict). J with F detects a ±4% focal change at k = 6 (pooled over sign and camera) in fewer than half as many splits as J with χ², on both detectors. Refuted if F keeps at least half on either. Milestone 4b found J already overstates the spread (rms z 0.46–0.84), and ν is 3–9 here, so F's threshold is several times χ²'s.
- H30 (a transferred threshold travels). J with E: S0 false alarms (k ≥ 4) ≤ 5% on each camera, on both detectors. Refuted if above 5% on any. Unlike milestone 4's borrowed κ, the jackknife already scales to each camera, so only the tail shape is borrowed.
- H31 (the limit is information). With the oracle threshold O, a ±2% focal change at k = 6 (pooled over sign and camera) is detected in fewer than 25% of splits on both detectors. Refuted if 25% or more on either.
- H32 (testing f alone buys power). z<sub>f</sub>² against χ²<sub>1</sub> detects a ±2% focal change at k = 6 in more splits than W against χ²<sub>3</sub>, on both detectors, with S0 false alarms (k ≥ 4, pooled) ≤ 3%. Refuted if either part fails on either detector.
- Reported, unscored. The full power table (±0.5–4% focal, 2–10 px principal point, the swap) for every reference; harmful adoptions under milestone 4's rule with each reference; the threshold values (median ν, F and E thresholds).
- Caveat carried over. The splits reuse 13 views, so the 95% Wilson intervals are too narrow, and an empirical 99th percentile from 180 statistics rests on about two of them.
2026-11-12 · Milestone 4b result: a jackknife over views makes the update test quiet with no borrowed factor (1.7–2.5% false alarms), but it cannot see a 2% focal change; SB corners lift its swap detection from 30% to 75%
- Runs. 2,145 distinct view sets per camera and detector (old and new sets plus every leave-one-view-out subset needed for the jackknife), 8,580 solves, 4 not converged (all 3-view sets; a jackknife drops a failed delete-one solve). The other camera's nested κ recomputed with H8's formula: 11×11 corners 2.67 (left) / 1.89 (right), as milestone 3; SB corners 3.47 / 2.74. All scored on k ≥ 4 (360 decisions per detector and covariance); 95% Wilson intervals.
- H23 held. Reported covariance (R), SB corners: S0 false alarms 57.8% (53–63), against 74.2% (69–78) with 11×11 corners. Better corners cut false alarms by a fifth, not to the nominal 1%.
- H24 held. Sandwich clustered by view (V), with G/(G − 1): 37.8% (11×11) and 36.1% (SB). 4–7 clusters per set are too few for a sandwich.
- H25 held. Jackknife by view (J): 2.5% (1.3–4.7; left 2.8%, right 2.2%) on 11×11 corners and 1.7% (0.8–3.6; left 0.6%, right 2.8%) on SB, with no factor taken from anywhere. The per-parameter rms z on S0 is 0.46–0.84, so J overstates the spread on average; that it still fires above 1% says the tails are heavier than χ²<sub>3</sub>.
- H26 held. The real unit swap (k ≥ 4, both directions): J 30.3% (11×11) and 74.7% (SB), against the transferred-κ test (T) at 41.7% and 93.1%. R and V detect it in 92–100%, but at their false-alarm rates that is not a test.
- H27 held. Harmful adoptions with J, all scenarios, k ≥ 4: 0.47% (11×11) and 1.54% (SB) of decisions. With the gate also on the jackknife σ<sub>f</sub> (unscored), 0.04% / 0.13%, but the rule then adopts in only 0.8–1.0% of decisions (13% with the reported-σ gate on SB).
- Unscored: power with J, 6 new views. ±1% focal change detected in 2–4% of splits (both detectors); ±2% in 7–12%; +4% in 46% / 49%, −4% in 41% / 62%. A 10 px principal-point shift: 6% / 12%. With SB corners, T keeps false alarms at 0.6% pooled (left 1.1%, right 0.0%) and detects ±4% in 70–88% and the swap in 93%: on the cleaner corners the borrowed factor no longer over-alarms on the left camera (11.7% with 11×11 corners).
- Reading. Five of five predictions held. An honest covariance from the views themselves exists and needs no outside calibration of the test, but the honest answer is that 4–6 views of this rig cannot tell a 2% focal change from the set-to-set offset. The corner localiser matters as much as the covariance: SB corners take the jackknife test from missing most real swaps to catching three in four.
- Caveat. As in milestone 4, the splits reuse 13 views, so the intervals are too narrow; J needs k + 7 extra solves per decision (about 1–3 s each here).
2026-11-12 · Milestone 4b planned: does a view-level covariance (jackknife or clustered sandwich) make the update test honest, and do SB corners change it?
Written before any milestone-4b solve was run. Milestone 4 left the χ² update test firing on 72.5% of unchanged splits with the reported covariance, and a scalar κ borrowed from the other camera (H19) under- or over-corrected per camera (10.8% left, 2.1% right). Milestone 3 found the leave-one-view-out jackknife σ of f conservative (rms z 0.65–0.70) where the reported σ is overconfident (≈ 2×). Here the scalar κ is replaced by a covariance each calibration computes from its own views, and the whole test is rerun on milestone 6's findChessboardCornersSB corners.
- Same splits. Milestone 4's 60 splits per camera and k ∈ {3, 4, 5, 6} (old set of 7 views, new set of k from the other 6, seeds 4000 + 100k + i), the same disturbances S0–S3, the same rule (χ²<sub>3</sub> on f, c<sub>x</sub>, c<sub>y</sub> at 1%, adopt iff detected and reported σ<sub>f,new</sub>/f ≤ 0.5%), for two detectors:
cornerSubPix 11×11 (milestones 1–5) and SB. - Four covariances for Σ<sub>old</sub> + Σ<sub>new</sub>. (R) reported (κ = 1); (T) reported × the other camera's nested rms z for that detector (milestone 3 for 11×11; milestone 6's SB subsets for SB, computed with H8's formula); (V) the 3×3 sandwich clustered by view, with the small-sample factor G/(G − 1); (J) the 3×3 leave-one-view-out jackknife, (m − 1)/m Σ(θ<sub>i</sub> − θ̄)(θ<sub>i</sub> − θ̄)ᵀ, each delete-one set re-solved cold from reference views 0 and n // 2. J needs k ≥ 4 (a 2-view delete-one set is not identifiable), so J's hypotheses are scored on k ≥ 4; V and R also on k ≥ 4 for the comparison. Under S1/S2 every covariance is scaled by s² as in milestone 4.
- H23 (better corners do not fix the false alarms). With R and SB corners, S0 false alarms (k ≥ 4, pooled) ≥ 50%. Refuted if below 50%. Milestone 6 found SB corners shrink σ more than the error.
- H24 (few clusters bias the sandwich down). With V, S0 false alarms (k ≥ 4, pooled) > 5% on both detectors. Refuted if ≤ 5% on either.
- H25 (the jackknife makes the test honest). With J, S0 false alarms (k ≥ 4) ≤ 3% pooled and ≤ 5% on each camera, on both detectors. Refuted otherwise.
- H26 (the jackknife pays in power). With J, the real unit swap S3 (k ≥ 4, pooled over directions) is detected in fewer splits than with T, on both detectors. Refuted if J detects at least as many on either detector.
- H27 (no harmful adoption with J). With J, over all scenarios with k ≥ 4, adoptions whose f error exceeds the old calibration's by more than 0.5% of f ≤ 2% of decisions, on both detectors. Refuted if more on either.
- Reported, unscored. The full power table (focal changes ±0.5–4%, principal-point shifts 2–10 px) for each covariance and detector; the gate applied with the jackknife σ<sub>f</sub> instead of the reported one; the per-parameter rms z on S0.
2026-11-12 · Milestone 4 result: the update test cries wolf on the raw covariance and goes deaf when inflated; each view set carries its own offset
- Runs. 908 distinct view sets (old sets of 7, new sets of 3–6, all-view), all converged. The disturbance transform was checked by re-solving disturbed synthetic corners (match to 1e-4 px,
tests/test_selfcal.py). - H18 held. Unchanged camera, κ = 1: the 3-parameter test fires on 72.5% of splits (95% 68–76; left 68%, right 77%); the f-only test on 22–30%.
- H19 refuted. With the other camera's factor (1.89 left, 2.67 right): 6.5% pooled (4.6–9.0), 10.8% left, 2.1% right.
- H20 refuted. The real unit swap, k ≥ 4: detected in 77% (new views from the left camera) and 30% (from the right). The f-only test almost never sees it (0–5%): the swap shows in the principal point.
- H21 refuted. A 1% focal change, k ≥ 5: 11.7% (−1%) and 10.0% (+1%). 4% is detected in 62–75% at k = 5–6.
- H22 refuted. Harmful adoptions (new f error 0.5% of f worse than the old) 3.5% of all decisions (3.2–3.9); 5.6% with κ = 1, 1.4% with the transferred factor. Flag rate 37% / 7%.
- Unscored diagnostic: why. Old-vs-new rms z for (f, c<sub>x</sub>, c<sub>y</sub>): 2.45 / 2.35 / 2.94 (left), 2.33 / 4.56 / 2.51 (right). The right camera's c<sub>x</sub> stays at 4.3–4.8 from 3 to 6 new views: an offset per view set, not noise. The in-sample factor for exactly 1% false alarms is 2.69 / 4.80 (3-parameter) and 2.24 / 2.69 (f only); at it, a 2% focal change with 6 views is detected in 32–38% (left) and 0–2% (right) of splits, 4% in 90–97% and 52–72%.
- Caveat. 60 splits of 13 views reuse views heavily, so the intervals, which assume independent splits, are too narrow. Corners are milestones 1–5's 11×11
cornerSubPix; the next step reruns on SB corners.
2026-11-12 · Milestone 4 planned: the accept/reject test on disjoint views, a real camera swap and exact disturbances
Written before any milestone-4 experiment was run. Proposal milestone 4: turn milestones 1–3 into the decision a robot makes, and measure its false-alarm and false-adopt rates.
- Old and new from disjoint views. Milestones 2–3 scored subsets against an all-view solution that contains them (hence the nested variance). A deployed camera's old calibration comes from other images, so here each camera's 13 views are split: an "old" set of 7 views, and a "new" set of k ∈ {3, 4, 5, 6} views drawn from the other 6. 60 splits per camera and k (fixed seeds), both sets solved with the milestone-2 solver (k1 model). The two estimates are then independent apart from the shared board, and the test covariance is Σ<sub>old</sub> + Σ<sub>new</sub> (full 3×3 over f, c<sub>x</sub>, c<sub>y</sub>).
- Changes. (S0) none. (S1) focal change: the new set's corners scaled about the image centre by s = 1 ± 0.5, 1, 2, 4%. (S2) principal-point shift of 2, 5, 10 px in u. Both are exact transforms of the pinhole-plus-k1 model, so the new solution is the transformed undisturbed one (checked by re-solving in a test) and the real detector noise is kept. (S3) a real change: old from one camera, new from the other (the stereo pair is two units of the same model; all-view f differs by 0.67%, c<sub>x</sub> by 17 px, c<sub>y</sub> by 10 px). Truth: the camera's all-view target-free solution, transformed for S1/S2, the other camera's for S3.
- The rule. χ² = Δᵀ(κ²(Σ<sub>old</sub> + Σ<sub>new</sub>))⁻¹Δ on (f, c<sub>x</sub>, c<sub>y</sub>). Change detected iff χ² > 11.34 (χ²<sub>3</sub>, 1%). Detected and σ<sub>f,new</sub>/f ≤ 0.5% (milestone 2's gate): adopt the new calibration. Detected and imprecise: flag for a target recalibration. Not detected: keep the old one. κ = 1 (reported covariance) and κ = the other camera's milestone-3 nested rms z (1.89 for the left camera, 2.67 for the right), so no factor is fitted on the camera it is scored on.
- H18 (false alarms). On S0 with κ = 1 the test fires on more than 5% of splits (nominal 1%). Refuted if ≤ 5%.
- H19 (the transferred κ). On S0 with transferred κ, false alarms ≤ 3% pooled and ≤ 5% on each camera. Refuted otherwise.
- H20 (the real swap). With transferred κ, S3 is detected in ≥ 90% of splits with k ≥ 4, in both directions. Refuted if below on either.
- H21 (power for a 1% focal change). With transferred κ, s = 1.01 and 0.99 are detected in ≥ 80% of splits with k ≥ 5. Refuted below. The power curve over all s, shifts and k is reported.
- H22 (no harmful adoption). Over all scenarios, k and both κ, the rule adopts an update whose f error against truth exceeds the old calibration's by more than 0.5% of f in ≤ 2% of splits. Refuted if more. Also reported: the flag rate, and how often the decision leaves an f error ≥ 1%.
- Ablation, unscored. The f-only test (1 degree of freedom) and the 4-parameter test with k1.
2026-11-11 · Milestone 6 result: the one-axis pattern was the corner localiser; it also made most of milestone 5's c<sub>y</sub> shift; better corners make σ less honest
- Localiser. Corners from
cornerSubPix with an 11×11 half-window (milestones 1–5) against a 5×5 window and findChessboardCornersSB (CALIB_CB_ACCURACY), matched one-to-one in all 26 images. Interior median gap 0.07 px (11 vs 5), but 26 corners differ by > 1 px (max 6.4 px), every one on the outer ring of the grid, where the 23 px box runs off the pattern. SB vs 11×11: 52 over 1 px, max 7.7, all on the ring; SB vs 5×5: max 1.7 px. - H14 refuted. SB along-column correlation 0.32 (left) / 0.56 (right), not < 0.2. The one-directional structure goes (along-row correlation rises to 0.23–0.34; per-(view, column) variance share 0.65 / 0.51 → 0.41 / 0.43, null 0.17), leaving a smoother 2-D correlation no localiser removes.
- H15 held. 91% / 78% of the per-(view, column) mean residual is perpendicular to the column, with the old corners (57–67% with the new ones).
- H16 refuted. (view, column)-clustered sandwich, original corners, random 4–12 views against all-view Zhang: rms z 2.44 / 1.72 (reported σ 2.21 / 1.63). Synthetic check (in the tests): with pure per-column offsets it rises above the reported σ but still covers only part of the true spread (4.3 against 5.8 px; reported 3.6).
- H17 refuted, opposite direction. Nested rms z with SB corners 3.12 (95% 2.88–3.37), 5×5 3.45, against 2.28 with the old corners; median |z| 1.97 / 1.70 / 1.40. σ halves (median 1.44 → 0.87 / 0.73 px) while the error against each detector's own all-view fit falls only from 0.24% to 0.23% / 0.15%. Fewer subsets drop from the nested test with the new corners (7–11 vs 58).
- Unplanned, corrects milestone 5. The flat-board c<sub>y</sub> shift is 8.7 px (old corners), 4.5 px (5×5) and 1.2 px, 1.8σ (SB) on the left camera; 5.3 / 3.8 / 1.4 px on the right. The bend survives every detector (flatness rms 0.0042–0.0058 against 0.0077; left/right height correlation 0.89–0.90). Milestone 5's "5–10 px" was mostly misplaced outer corners, which a free-board fit absorbs as bend.
- Unplanned. The localiser alone moves flat-board Zhang's f by 4.8 px (left) and 7.2 px (right), 0.9–1.3%, and the target-free f by 1.8 / 4.2 px, while every fit reports σ ≤ 1.2 px. No ground truth says which localiser is right.
- Deviation. The SB matching tolerance was first 3 px and failed on images with 4–8 px outliers (which were the finding); raised to 10 px (half the smallest corner spacing), with the one-to-one check kept. All 1,680 subset solves ran after that; 1,679 converged.
2026-11-11 · Milestone 6 planned: where does the column-wise residual correlation come from, and does a better corner detector or a column-clustered σ fix the 2×?
Written before any milestone-6 experiment was run (only the corner spacing and the detection rate of the second detector were checked: 25–41 px between neighbours, and findChessboardCornersSB finds the board in 26 of 26 images).
- What milestone 5 left. In the target-free k1 residuals on all 13 views, corners that are neighbours along a board column correlate at 0.56 / 0.44 (left / right), and corners two apart at 0.73 / 0.65, while neighbours along a row are at about 0. A correlation that rises with distance looks like an offset shared by each column of each image, not a local detector error. Free 3-D points absorb any fixed print error of the board, so whatever is left must change from image to image.
- Descriptive, not a test. The share of residual variance explained by a per-(view, board column) mean and by a per-(view, board row) mean, against the null for independent noise (1/6 and 1/9); and how much of the column-mean residual lies perpendicular to the column's line in the image, against along it.
- H14 (the detector is the cause). Detect the same corners with a different localiser, OpenCV's
findChessboardCornersSB with CALIB_CB_ACCURACY (Duda and Frese, 2018), and, as an ablation, cornerSubPix with a 5×5 half-window instead of 11×11. Predicted: with SB corners the adjacent-along-column correlation of the all-view target-free residuals falls below 0.2 on both cameras. Refuted if it stays at 0.2 or more on either, which puts the cause upstream of the localiser (the image or the board's mounting). - H15 (the column offsets are sideways). The per-(view, column) mean residual is predicted to lie mostly perpendicular to the column's image direction: ≥ 70% of its energy on both cameras (isotropic would be 50%). A sideways shift of a whole column points at the localisation of that column's edge line; a shift along it does not.
- H16 (a column-clustered σ). The sandwich σ of f clustered by (view, board column) — 9k clusters, which keeps the within-column correlation and drops only the cross-column part of the view clustering — on milestone 2's random 4–12-view subsets, original corners, scored against all-view Zhang as in milestone 3. Predicted to lie in [0.7, 1.4] on each camera. Refuted otherwise. Also reported, unscored: the (view, row) clustering and the nested test against the target-free all-view solution.
- H17 (a better detector shrinks the 2×). Re-solve milestone 2's 560 subsets (same views and reference views) with SB corners and the k1 model, and score the nested self-consistency test of H8 against the SB all-view target-free solution. Predicted: rms z ≤ 1.5 (against 2.28 with
cornerSubPix). Refuted if > 1.5. Also reported: the all-view intrinsics with each detector, and whether c<sub>y</sub> and f move.
2026-11-11 · Milestone 5 result: the bent board moves c<sub>y</sub> by 5–9 px; adding k2 does not make σ honest
- Released-object Zhang is target-free BA.
calibrateCameraROExtended and our target-free solver return the same f, c<sub>y</sub>, σ, board shape (height correlation 1.000) and residuals to 1e-6 px on both cameras and both lens models. With the corners free, the only extra information the board gives is its metric scale, which carries none about the intrinsics. So H11's test (gap < 2σ) held trivially (gap 4e-6σ), and the evidence for the bend is the flat-vs-released comparison: the only difference between the two Zhang fits is the flat, evenly spaced board, and it moves c<sub>y</sub> by −8.7 px (left) and −5.3 px (right) and f by +0.79 / +1.28 px (k1 model). - H12 held, more strongly than predicted. With k1 + k2, nested self-consistency rms z is 3.01 (95% 2.05–4.04; left 3.85, right 2.04), against 2.28 with k1 alone. Fragile: two left-camera subsets of 4–5 views (z −28, −23) carry it, where σ<sub>sub</sub> barely exceeds σ<sub>all</sub>; without the top five |z| it is 2.00 (k1: 1.94), median |z| 1.31 (k1: 1.40). Either way k2 is not the cause. All 560 solves converged (at most 180 evaluations); k2 = 0.143 ± 0.018 (left) and 0.091 ± 0.007 (right), clearly non-zero.
- H13 held. Adjacent-corner residual correlation with k1 + k2: 0.31 (left) and 0.24 (right); flat-board Zhang 0.36 / 0.32. The structure is one-directional: along the board's columns 0.44–0.56 (0.76 at two corners apart), along its rows −0.02 to 0.02. A lens term would act radially, not along one board axis; the cause (print, detector, the board's mounting) is not established.
- Unscored. Against the released-object reference, rms z is 1.66 (k1) and 1.79 (k2): adding k2 shrinks σ<sub>f</sub> by a median 5% and leaves the median error unchanged (0.24%).
- Correction to milestone 1. The reported reprojection "0.42 vs 0.26 px" compared OpenCV's per-point rms (√(u² + v²)) with our per-coordinate rms. On one definition, per coordinate: flat Zhang 0.30 / 0.34 px, target-free 0.26 / 0.29 px, 13–15% lower, not about 40%. Fixed in the README and card.
2026-11-11 · Milestone 5 planned: does the bent board explain the principal-point gap, and does an unmodelled k2 explain the 2× σ?
Written before any milestone-5 experiment was run. Proposal milestone 5 (unmodelled lens terms on real data), taken before 3b and 4 because it needs no new data and milestones 1 and 3 left two named suspects untested: the bent board (for the c<sub>y</sub> gap) and the k1-only lens (for the overconfidence).
- Bend-aware reference. Zhang with a released object (
cv2.calibrateCameraROExtended, Strobl and Hirzinger 2011): the board's 3-D corners are refined with the camera, two corners fixed (OpenCV's iFixedPoint = 8, the last corner of the first row). Same model (f, c<sub>x</sub>, c<sub>y</sub>, k1; k2 as ablation), all 13 views. - H11 (the bend explains c<sub>y</sub>). With the released-object reference, |c<sub>y,RO</sub> − c<sub>y,target-free</sub>| < 2σ on both cameras (σ: both solvers' σ in quadrature), against 5.3σ and 2.9σ with the flat board. Refuted if ≥ 2σ on either camera.
- H12 (k2 is not the cause of the 2×). Target-free BA with k1 and k2 on milestone 2's 560 subsets (same views and reference views). The nested self-consistency rms z of milestone 3 (H8; against the k1+k2 all-view solution) stays > 1.5. Refuted if ≤ 1.5, which would make the unmodelled k2 the main cause.
- H13 (structured residuals survive k2). On the all-view fits, the correlation of the residual vectors of horizontally and vertically adjacent corners in the same image is ≥ 0.1 with k1 and k2 (correlated detector error, not lens model). Refuted if it falls below 0.1 with k2.
- Also reported, not scored: rms z of the k1+k2 subsets against the released-object Zhang f, and the f shift from adding k2 in each solver.
2026-11-11 · Milestone 3 result: the overconfidence is real; the jackknife σ is honest but a worse gate
- Re-solves identical. All 560 subsets converged with the same f as milestone 2; 3,473 of 3,480 leave-one-view-out solves converged on tolerance. Median 2.7 s per subset with its jackknife.
- H8 held. Against the target-free all-view f, nested variance: rms z 2.28 (95% 1.93–2.73; left 2.67, right 1.89; 58 of 360 subsets dropped with σ_sub ≤ σ_all). The 2× was not the Zhang reference. Jackknife σ in the same test: 0.85.
- H9 refuted, on the safe side. Jackknife rms z 0.703 (left) and 0.651 (right; 95% 0.56–0.74), 2σ coverage 97% / 99%; near-critical 4–6 views 1.01 (reported σ 2.34). By size: 0.33 at 4 views (σ_jack a median 15× the reported σ), 0.55–0.94 from 5 to 11. The view-clustered sandwich: 1.76 / 1.59, and 2.23 at 4–6 views as predicted; corner-clustered and HC0 sandwiches are no better than the reported σ.
- H10 refuted. AUC for good (≤ 1%) against bad (> 2%) updates, 4–12 views: reported σ 0.91, jackknife 0.75 (difference −0.15, 95% −0.25 to −0.07), sandwich by view 0.76, by corner 0.87, HC0 0.89. Only 22 bad updates.
- H4 second test refuted. σ_jack/f ≤ 0.5% rejects 22/22 bad but accepts 19% of good (0% near-critical). Transfer: no σ's fitted threshold reaches 90/90 on the other camera; the reported σ comes closest (76/79 and 91/88).
- Reading. The jackknife's upward bias (Efron–Stein) is largest where leave-one-out leaves 3 views. Honesty and ranking power pull apart: next gate ranks by reported σ and spread, and vetoes by the jackknife.
2026-11-11 · Milestone 3 planned: is the 2× overconfidence real, and does a cluster-robust σ fix it?
Written before any milestone-3 experiment was run. The proposal's milestone 3 (natural features on a second data set) moves to 3b; this step answers the question milestone 2 left open, on the same 560 subsets, with no new data.
- Why σ could be too small. The Gauss–Newton σ assumes independent, equal-variance pixel residuals and a correct model. Corner-detector errors are correlated within an image (blur, lighting, the detector's shared sub-pixel window) and the k1-only lens leaves structured residuals. Both make the reported σ too small by a factor that differs between subsets, which a single inflation factor cannot follow (milestone 2's cross-camera κ).
- Robust σ, two estimators, no fitted constants. (a) The cluster-robust sandwich, clustered by view: var f = Σ_views ((J b)·r)² over each view's residuals, with b = (JᵀJ)⁻¹e_f, times the small-sample factor G/(G − 1). (b) The leave-one-view-out jackknife: re-solve the subset k times without one view each (cold start, same solver), var f = (k − 1)/k Σ (f₍ᵢ₎ − f̄)². Ablation: the sandwich clustered by corner (54 clusters) and unclustered (HC0).
- H8 (the reference is not the cause). Zhang's all-view f is itself 0.8 / 1.3 px from the target-free all-view f, so part of milestone 2's z could be reference mismatch. Self-consistency test against the target-free all-view solution, with the nested-estimator variance: z_n = (f_sub − f_all) / √(σ_sub² − σ_all²), random subsets of 4–12 views (subsets with σ_sub ≤ σ_all are counted and dropped). Predicted: rms z_n > 1.5 (the overconfidence is real). Refuted if ≤ 1.5.
- H9 (robust σ is honest). With σ from the jackknife, rms z against all-view Zhang over random 4–12-view subsets lies in [0.7, 1.4] on each camera, with no fitted factor. Scored separately for the view-clustered sandwich; predicted to hold for the jackknife and to fail (rms z > 1.4) for the sandwich at 4–6 views, where few clusters bias it down.
- H10 (robust σ ranks better). As a gate statistic for good (|Δf|/f ≤ 1%) against bad (> 2%) updates, pooled over all subsets of 4–12 views and both cameras, the jackknife σ/f has a higher ROC AUC than the reported σ/f. Refuted if it is equal or lower (paired bootstrap over subsets, 95% interval of the AUC difference).
- H4, second test. Accept iff σ_jack / f ≤ 0.5% (the same threshold as milestone 2); target 90/90. Also scored on the near-critical subsets alone. And the transfer test from milestone 1's next steps: the threshold on each σ that maximises min(accept good, reject bad) on one camera, scored on the other.
2026-11-11 · Milestone 2 result: geometry predicts the error, σ stays 2× small, and Zhang's σ is far worse near the critical set
- Solver fixed. The cap was not only the gauge: holding view
ref's pose fixed left the scale free (the initialisation puts that camera at the world origin, so its translation is zero), and scipy's trf/lsmr crawled even once the scale was pinned too (corner 0's depth). Levenberg–Marquardt on a dense analytic Jacobian (checked against finite differences, max error 1e-6 relative) converges in 21–23 evaluations on all 13 views and in at most 177 on any subset: 560 of 560 solves converged on tolerance. All-view f moved by 0.0004 px (534.8239 / 538.4302). - H2 rerun held. Same 360 subsets of 4–12 views: rms z 1.945 (95% 1.75–2.12) against milestone 1's 1.944; median f change 0.002 px, largest 0.66 px. The cap was not the cause of the 2× overconfidence.
- H6 held. Spearman ρ between view spread and |Δf|/f: −0.46 (left, n 240) and −0.58 (right, n 240), p < 1e-13. The reported σ also tracks the error (ρ +0.39 / +0.45).
- H7 refuted. On the 160 near-critical subsets (spread 7–36°) rms z is 2.41 (95% 2.09–2.81); 20% have |z| > 3. Unplanned but striking: Zhang's σ on the same subsets is far less honest, rms z 5.83, and its median error is larger than the target-free one at 4–6 views (1.50 / 1.15 / 1.08% against 1.09 / 0.99 / 0.82%).
- H4 first test refuted. The 0.5% precision gate rejected 92% (56 of 61) of bad updates (> 2% off) but accepted only 84% (340 of 407) of good ones (≤ 1%). The five bad accepts were 2.0–2.9% off with σ/f 0.45–0.48%, just under the gate. Doubling σ (unplanned) rejects every bad update but accepts only 37% of good ones.
- Unplanned: σ inflation does not transfer cleanly between cameras. κ = rms z fitted on left (2.21) gives right-camera rms z 0.74 (over-cautious); κ from right (1.63) gives left rms z 1.35 and 2σ coverage 85% (random) and 82% (near-critical). Median error / σ by σ bin is 1.0–1.8, so the overconfidence sits in the tails more than in the typical subset.
- No rotation-only subsets. The milestone plan named rotation-only motions; this data set has a moving board and no such views, so near-critical here means near-parallel board planes only.
2026-11-11 · Milestone 2 planned: does the view geometry predict the error, and can σ gate an update?
Written before any milestone-2 experiment was run.
- Solver fix first. Milestone 1's solves all stopped on the 1,000-evaluation cap. Fix the gauge by holding view 0's pose (rotation and translation) constant: that removes all 7 similarity directions (a fixed non-zero translation pins the scale), so the normal matrix is invertible and the intrinsics' covariance is the plain inverse, equal to the gauge-invariant pseudo-inverse. Raise the cap and record convergence status for every solve.
- H2 rerun (was the cap an artefact?). Rerun milestone 1's 400 random subsets (same seeds) with the fixed solver. Predicted: rms z for 4–12 views stays within ±0.2 of 1.94 (the all-view f already agreed across starts to 0.001 px). Refuted if it moves by more than 0.2.
- View spread. For each view, the board normal in the camera frame from the all-view Zhang poses. A subset's spread is the largest angle between two of its normals. All parallel normals is the planar critical case (Sturm and Maybank 1999): f is then unidentifiable.
- Near-critical subsets. For sizes 3, 4, 5 and 6, enumerate every subset of the 13 views and keep the 20 with the smallest spread, per camera (160 solves), alongside the 400 random ones.
- H6 (geometry predicts error). Pooled over random and near-critical subsets of 4–12 views, Spearman ρ between spread and |Δf|/f (against all-view Zhang) is ≤ −0.3 on each camera. Refuted if ρ > −0.3 on either.
- H7 (σ flags near-critical subsets). On the near-critical subsets, rms z ≤ 2, i.e. σ grows as fast as the error. Refuted if rms z > 2.
- H4, first test (a precision gate). Accept an update iff σ_f / f ≤ 0.5%. A good update has |Δf|/f ≤ 1%, a bad one > 2% (between is not scored). Pooled over all subsets and both cameras, the gate should reject ≥ 90% of bad updates and accept ≥ 90% of good ones. Refuted if either rate is below 90%. The proposal's χ² test against an old calibration needs a disturbed camera, which this data set does not have; that test waits for milestone 4.
2026-11-11 · Milestone 1 result: target-free BA matches Zhang to 0.2%, the board bends, σ is 2× optimistic
- H1 held. All 13 views: target-free f 534.82 / 538.43 px against Zhang's 535.62 / 539.71 (left / right), gaps −0.15% / −0.24%, under 1σ; the same f from 0.7× and 1.3× starts.
- H2 held narrowly. Subsets of 4–12 views (360 solves): rms z 1.94, 95% bootstrap 1.75–2.12. Ideal is 1, so σ is about twice too small; worst at 3–7 views (2.1–2.8). Zhang's own σ against the same reference: 2.16. Not planned: subsets of 3 views (rms z 2.79, worst error 20%).
- H3 refuted. σ ratio target-free / Zhang 1.12 at 13 views (predicted ≥ 2); 2.9 at 3 views, falling with views as predicted.
- Unplanned finding: the board is not flat. The free corners leave their best plane by 0.0077 square widths rms (max 0.024) on both cameras, and the two cameras' out-of-plane maps correlate at 0.903 (0.909 with k2). Target-free reprojection rms 0.26 / 0.29 px against Zhang's 0.42 / 0.49. The principal point's c_y differs by 8.7 px (5.3σ) and 5.2 px (2.9σ); the bend is the likely cause, not yet tested.
- Deviation from the plan. Zhang's σ_f came out as 0 at first: with a fixed aspect ratio OpenCV reports it in the f_y slot; fixed before any analysis. Sizes 3–12 were run (plan said 4–12); H2 is scored on 4–12 as planned.
- Problem found. Every solve stopped on the 1,000-evaluation cap (the similarity gauge slows trust-region). For all-view fits f agrees across starts to 0.001 px; subsets were not checked. Milestone 2 should fix the gauge and use an analytic Jacobian.
2026-11-11 · Milestone 1 planned: target-free self-calibration against Zhang on a real stereo pair
Written before any experiment was run.
- Data. OpenCV's
samples/data stereo calibration set, left01..14.jpg and right01..14.jpg (index 10 absent in both), 640×480, a 9×6-inner-corner checkerboard, Apache-2.0. Two real cameras, 13 views each. - Reference. Zhang calibration with the board's known geometry (
cv2.calibrateCamera, model fx = fy = f, cx, cy, k1, k2) on all 13 views of each camera; interval by bootstrap over views. - Target-free solver. The same detected corners, but the 54 board points become free 3-D unknowns (no planarity, no spacing): initial focal guessed 30% wrong, two-view initialisation from the homography between images (no board metric), PnP for the rest, then bundle adjustment over poses, points and intrinsics (f, cx, cy, k1). Covariance from the Gauss–Newton normal matrix with the gauge fixed, scaled by the residual variance.
- H1. On all 13 views, target-free f lies within 2% of Zhang's f for both cameras. Refuted if either is further.
- H2 (the solver knows when it is wrong). Over random view subsets of size 4 to 12, the z-score (target-free f − all-view Zhang f) / σ_f(target-free) has rms ≤ 2. Refuted if rms > 2, i.e. the covariance is overconfident.
- H3 (price of no target). Target-free σ_f is at least twice Zhang's on the same views, and the ratio falls as views are added. Refuted if the ratio is below 2 at 13 views.
Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.