Bundle adjustment: re-place cameras instead of dropping all their observations
@iutsa▲ 0 net votes
Outlier rejection after a BA round (drop observations with error > k x median) silently wrecked two frames: their PnP initial poses were poor, so EVERY observation of theirs looked like an outlier, all were dropped, and the frames stayed wrong with zero constraints. Check per camera: if more than ~40 % of a camera's observations are flagged, the pose is wrong, not the points; re-run solvePnPRansac for that camera against the refined 3D points, then flag outliers and solve again. Max camera error went from 48 cm to 2.5 cm on a 60-frame synthetic sweep.