RA-LoFTR: Rotation-Equivariant Annular Convolution for Robust Detector-Free Matching
TL;DR
RA-LoFTR is proposed, which enhances LoFTR with Rotational Coordinate Convolution and Adaptive Fusion Weighting and dynamically fuses RCC-derived local structural cues with Transformer-based global positional information, and MAGSAC++ is further used as a geometric verification step for outlier rejection.
Abstract
Image feature matching is a cornerstone of computer vision, yet robust correspondence estimation under rotation, low texture, and illumination variation remains challenging. Detector-free methods such as LoFTR reduce the dependence on repeatable keypoints, but their standard convolutional backbones are still sensitive to orientation changes and weak local structures. To address these limitations, we propose RA-LoFTR, which enhances LoFTR with Rotational Coordinate Convolution (RCC) and Adaptive Fusion Weighting (AFW). RCC partitions feature neighborhoods into concentric annular regions and aligns dominant orientations through channel-wise cyclic shifts, transforming rotation handling from discrete angle classification into channel-phase alignment. AFW dynamically fuses RCC-derived local structural cues with Transformer-based global positional information, and MAGSAC++ is further used as a geometric verification step for outlier rejection. On MegaDepth, RA-LoFTR achieves pose-estimation AUCs of 37.4, 53.6, and 66.0 at 5°, 10°, and 20°, improving over LoFTR by 1.1, 1.7, and 1.3 points, respectively. On HPatches, it improves homography AUCs over LoFTR by 2.2, 1.5, and 1.9 points at 3, 5, and 10 pixels, respectively. Robustness experiments under low-light, motion blur, Gaussian noise, and low-texture conditions further validate the proposed annular convolution design.