XGMR: Detector-Free RGB-Thermal Image Registration with Geometric Inductive Bias

Reference implementation of XGMR, a detector-free, self-supervised framework for RGB-thermal image registration with geometric inductive bias (epipolar-biased cross-attention plus an equivariance-based self-supervised loss). This repository contains the source code, configuration files, and the exact train/validation/test split lists for the three benchmark datasets (NII-CU MAPD, LLVIP, MSRS) used in the paper. Trained model checkpoints (XGMR_Full, three seeds, seven ablations, and a fine-tuned LoFTR baseline) are distributed as assets of the corresponding GitHub release. The benchmark images are third-party datasets and are not redistributed.