XoFTR++: dual-stage cross-modal enhancement for visible-thermal point correspondence matching

Visible-light-to-thermal-infrared feature matching is essential for multi-sensor fusion but remains challenging because of substantial modality discrepancies in image-formation mechanisms, texture statistics, and structural responses. Existing coarse-to-fine frameworks such as XoFTR improve cross-modal matching, yet their cross-modal interaction is concentrated primarily in the Transformer matching stage, leaving stage-wise error propagation insufficiently addressed. This study presents XoFTR++, a dual-stage cross-modal enhancement framework for visible–thermal correspondence matching. At the coarse stage, Cross-Modal Feature Enhancement (CMFE), Adaptive Feature Pyramid Fusion (AFPF), and Coarse-Level Transformer Enhancement (CLTE) are introduced to reduce early modality shifts, inject multi-scale semantic context into fine features, and improve coarse-candidate reliability. At the fine stage, Fine-Level Cross-Modal Attention Enhancement (FCAE) models local context, modality differences, and bidirectional window interactions to reduce local matching ambiguity before final similarity estimation. On METU-VisTIR under RANSAC = 1.5, XoFTR + + improves AUC@5/10/20 from 12.31 ± 0.62/28.03 ± 0.52/44.92 ± 0.49 to 15.07 ± 0.44/29.95 ± 1.32/47.14 ± 1.28 in Sunny–Cloudy scenes, and from 22.05 ± 1.72/39.11 ± 1.63/54.96 ± 1.59 to 23.28 ± 0.52/41.02 ± 1.04/57.13 ± 1.21 in Cloudy–Cloudy scenes, with a 10.29% parameter increase. Controlled ablations and stage-wise diagnostics support the complementary contributions of the four modules. A separate synthetic-homography evaluation on the public FusionDN RoadScene and MSRS datasets improves aggregate AUC@5/10/20 and reduces mean corner error, providing supplementary transfer evidence under the specified external protocol.