Tests whether feature correspondences between revisit frames agree with their ViPE-estimated camera geometry. This is a pose/feature consistency proxy, not an independent 3D ground-truth check.
For relative pose Tb−1Ta=(R,t) and nonzero baseline, use unit translation direction to avoid arbitrary monocular translation scale in the epipolar calculation:
SIFT matches are triangulated. Positive depth is required in both cameras; fpos is measured against all matches, so perfect reprojection behind cameras cannot earn full credit. Default pixel scales are τepi=4 px and τrep=8 px. Too few finite triangulations yield no measurement.
At zero/near-zero baseline, E is degenerate, not evidence of success. Use the pose-induced rotation homography H=KbRKa−1 and symmetric transfer error, with score exp(−e/τrep); epipolar error is not applicable. H is not fitted to the same matches.
Higher is better. The rollout averages measured pairs subject to pair-support requirements. A failed return gate does not block scoring; nearest-pose fallback pairs are used when strict revisits are absent, with gate status and pair source recorded. Missing measurement is distinct from a measured zero. The near-zero baseline tolerance is numerical, not physically calibrated; feature overlap and estimator reliability are limitations. Backend: sift_epipolar.
