Characterizing Correspondence Coverage Collapse in UAV RGB–TIR Registration

Registration quality in unmanned aerial vehicle (UAV) RGB and thermal infrared (TIR) imagery is often assessed using reprojection error, although low residuals at the correspondences used for estimation do not necessarily indicate reliable alignment across the common image footprint. This study investigates Correspondence Coverage Collapse as a diagnostic pattern in which low in-sample error coexists with concentrated correspondences, limited spatial support, and poorer held-out performance. Seventeen RGB–TIR image pairs were analyzed using a two-stage common-footprint normalization and a fixed full-inlier correspondence pool generated by SuperPoint–SuperGlue. Clustered, Random, and Balanced subsets were compared under equal correspondence budgets, with common held-out correspondences used for generalization assessment. At N=80, Clustered selection achieved the lowest selected-point RMSE but showed the highest held-out RMSE, lowest spatial coverage, and largest reference-grid deviation. Confidence-ranked selection showed a similar but weaker pattern at lower correspondence budgets. Continuous analyses further showed that lower coverage and greater distance-to-support were associated with higher held-out error and global transformation deviation. These results indicate that reprojection error alone does not characterize spatial support or held-out performance within the evaluated correspondence pool.