The paper introduces "consensus capture," a failure mode in reference-frame UAV video stabilization where coherently moving vehicles over urban intersections generate mutually consistent feature correspondences that bias robust homography estimates toward foreground motion, undetectable by standard inlier statistics. A two-part benchmark evaluates ten registration pipelines (AKAZE, RootSIFT, SIFT, BRISK, ORB, etc.) on 4K aerial footage. The static protocol uses exactly known homographies on site-disjoint data at matched resolution and feature budgets, showing tuned AKAZE, RootSIFT, SIFT, and BRISK leading; retuning matcher thresholds improves mean average accuracy (mAA) by at most 0.0062, while object masking costs under 0.005 mAA. The dynamic protocol tests offline and real-time budgets on hovers spanning varying foreground motion coherence, plus a stationary control, measuring background alignment, object position error, and the fraction of a moving group's displacement captured, against pseudo-ground-truth references audited with manual ground control points. Unmasked ORB suffers capture on 28 of 304 real-time pairs and 6 offline pairs; captured real-time estimates exhibit a median corner error of 18.8 pixels despite typical inlier counts and ratios. Object masking eliminates capture for all pipelines without altering the ranking of top performers. Code, data, and results are released.
Read original
huggingface/daily-papers