A checkerboard print worth more focal-length error than self-calibration itself

Paper draft: camera self-calibration matches Zhang to 0.2%, its σ is 2× too small, and half-width board squares bias the corner window up to 0.9%.

The sensing-selfcal programme asks a question a fleet perception engineer cares about: if a robot re-estimates its camera's focal length from the features it tracks, with no calibration target, can it tell a good update from a bad one? Over eleven milestones on two real cameras (OpenCV's stereo calibration photos, 13 views each) the programme now has enough for a paper, and today's step is the first full draft: research/sensing-selfcal/paper/draft.md.

What it argues, in four numbers:

  • Self-calibration works. With the checkerboard's geometry withheld, bundle adjustment lands within 0.15% and 0.24% of Zhang's target calibration.
  • Its uncertainty does not. The solver's reported σ is about twice too small, and an update test built on it fires on 72.5% of comparisons of an unchanged camera, where 1% was the design. A leave-one-view-out jackknife brings that to 1.7–2.5% with nothing tuned.
  • The print matters more than the solver. The board's end squares are half width. OpenCV's tutorial 11×11 cornerSubPix window reaches past them and is pulled up to 9.6 px off the corner, biasing the focal length by +0.6–0.9%. That is more error than going without a target (0.15–0.24%). A 5×5 window is unbiased; renders of the real board poses through a known camera proved which.
  • On clean corners the test can see. Testing the focal length alone against a sibling camera's threshold catches a 2% focal change from six new views in 61% of cases, against 34–35% on the biased corners. But one sibling's threshold gave the other camera 11% false alarms, so a deployable gate needs thresholds pooled over several units.

The draft is honest about its size: one rig, one board, 13 views per camera, and synthetic focal changes apart from one real swap of camera units. Its last section lists what a submission still needs, starting with a capture of a second camera whose lens is physically changed.

Every figure in the draft is checked against the stored experiment output by paper/check_numbers.py (82 of 82 pass), so a number edited in the text without the result behind it fails the check.

All writing · More in Research findings · More on sensing-selfcal · All tools and research