The sensing-selfcal programme is building a test a robot can run before it accepts a new camera calibration it worked out for itself. Last step, on unbiased corners, the focal-length test caught a 2% change in 61% of trials. The catch was its threshold. That came from a sibling camera, and it gave 11% false alarms on one of the two cameras. A fleet would rather have each camera set its own threshold.
The idea
A robot always keeps the views behind its current calibration. Split those 7 views into 4 + 3 (there are 35 ways), solve each half, and see how far the two focal lengths land apart. If each view adds its own small independent error, the spread shrinks as 1/n. That gives a prediction of how far an honest new calibration should land from the old one, and nothing is borrowed from another unit.
We wrote down five predictions first, then solved every 3- and 4-view subset of OpenCV's 13-view stereo set on three corner localisers (3,168 bundle adjustments) and scored 14,040 accept/reject decisions.
It is quiet, and deaf
The test almost never fires on an unchanged camera: 0.6% and 0.0% false alarms. It also almost never fires on a changed one. It catches a 2% focal change in 0.8% of splits, where the borrowed threshold caught 61%. The split-halves predict an old-to-new spread of 1.94% and 2.62% of f. The real spread is 0.67% and 0.50%, so they overstate it 2.9× and 5.2×. Three of the five predictions failed. The two that held (few false alarms, few harmful adoptions) held only because the test stays silent.
Why: small view sets are a different regime
After the failure we measured how the focal length's spread falls with the number of views, scaled so 1/n would give a flat line. Per view, on the unbiased corners:
| views in the set | 3 | 4 | 5 | 6 | 7 |
|---|---|---|---|---|---|
| left camera | 3.29% | 2.15% | 1.68% | 1.67% | 1.55% |
| right camera | 5.21% | 2.17% | 1.35% | 1.04% | 0.95% |
From 5 views up the line is roughly flat. Below that the error rises steeply: 3 or 4 boards seldom span enough orientations, and the solve sits near the critical configuration where focal length is barely observable. A 3-view half carries 2 to 5 times the per-view error of a 7-view set, so the threshold it implies is far too wide.
What it means for a fleet
A per-unit threshold is still the right goal, but its halves need at least 5 views each, so about 10 views behind the old calibration. Field robots log far more frames than that, so this asks for a bigger old set, not a new method. Until then a threshold pooled over sibling units is the only usable focal-length test. The limits still apply: one stereo rig and one board, 13 views shared by every split, and a post-hoc explanation that the next step should test on its own pre-registered data.
Code, data and every number: experiments/m4e_selfscale, and the paper draft's Section 7 (paper/draft.md).