Mismatch Removal for Wide-baseline Image Matching via Coherent Region-to-Region Correspondence

A novel wide-baseline image matching method that exploits coherent region correspondence is presented in this paper. The traditional matching algorithms solely rely on identifying the same local features between the two images to determine pixel-to-pixel matches. However, this kind of approach usually results in a high percentage of mismatches due to large viewpoint variations. In order to maximally reduce the mismatches, a bipartite graph model is used to establish coherent region-to-region matches between the segments of the two images. We further use this region correspondence to verify the correctness of each pixel-to-pixel match. Extensive experimental results have demonstrated that our proposed wide-baseline image matching method increases the percentage of correct matches compared with those methods that are based on the nearest neighbor matching criterion.