Support-Gated Dense Hypothesis Selection for Three-View Aerial Video Stitching

Three-view aerial video stitching requires spatial alignment and control of historical warp drift. We evaluate a target-training-free dense hypothesis-selection system using released LoFTR and PixelStitch components, observed third-view support, and a bounded historical proposal. Published synthetic depth and real laser geometry evaluate the actual backward sampler independently of the matcher. A locally frozen study uses six held-out SynDrone towns, three held-out ETH3D rig scenes and five estimator seeds. The guarded model improves synthetic success within 3 pixels from 70.576% to 71.033%, a 0.456-percentage-point difference with a descriptive town-bootstrap 95% interval of [0.187, 0.747] and an exact paired sign-randomization p = 0.03125. Real-rig success changes from 92.435% to 92.604%, but an adapted official PixelStitch control achieves 93.973%. Ablations attribute nearly all same-backbone gain to current spatial verification; the extra historical mechanism contributes little and does not consistently beat the spatial-only model. Complete fresh-matching inference and common rendering take 6.785 seconds per triplet, versus 4.530 for the pairwise control on the measured configuration. Two public multi-UAV clips demonstrate wide-angle output without independent dense accuracy truth. The evidence supports a modest spatial-selection improvement, not a general advantage from temporal memory, real-time operation or universal superiority.