MR-Matcher: A Multirouting Transformer-Based Network for Accurate Local Feature Matching

Local feature matching aims to identify pixel-wise correspondences between a pair of images and serves as an indispensable component of numerous visual applications. Recently, detector-free methods built upon transformers have acquired impressive results, surpassing typical convolutional neural network (CNN)-based methods by a significant margin. However, one concern about current transformer-based detector-free methods is their lack of concurrent integration of global, local, and multiscale features for feature aggregation, thus hampering the capability of the network to establish accurate and well-localized correspondences. Typically, local features can capture geometric structure information to bolster the representation ability of the network while multiscale features can promote the differentiation of regions or objects of diverse sizes to improve the robustness of the network to scale shifts. Specifically, we propose a multirouting transformer (MRFormer) block that consists of: 1) a multiscale representation (MR) layer to generate multiscale feature representations and 2) parallel local routing (L-Routing) and global routings (G-Routings) to acquire local and global feature representations based on 1), thus enabling multiscale, local, and global feature extraction at the same feature level. After obtaining such copious feature representations, we introduce a lightweight expert feature compensator (EFC) to effectively calibrate these feature representations from both local and global perspectives. Benefiting from the diverse feature representations, MR-Matcher exceeds the state-of-the-art approaches over multiple benchmarks.