Matching-to-Detecting: Establishing Dense and Reliable Correspondences Between Images

We present a novel view for local image feature matching, which is inspired by the difference between existing methods. Detector-based methods detect predefined keypoints in local regions, so that the stability and reliability of established matches are ensured. In contrast, detector-free methods usually directly match dense features and refine the filtered results, which can help generate more matches. In order to combine their advantages, we propose a novel Matching-to-Detecting (M2D) process for feature matching, in which we first perform global reasoning for patch-level matching and subsequently identify discriminative matches within local areas to obtain pixel-level matches. At the patch-level, dense matching provides our pipeline with the ability to find plenty of matches even in low-texture areas, while at the pixel-level, our method can be viewed as detecting from a matching perspective, so that the established matches have higher stability and reliability and are remarkable in local regions. Experimental results demonstrate that our method outperforms state-of-the-art methods by a significant margin in terms of matching accuracy and the number of matches. Moreover, the computational complexity of our model is quite low, making it more suitable for real-world applications.