Bipartite graph-based mismatch removal for wide-baseline image matching

The conventional wide-baseline image matching aims to establish point-to-point correspondence pairs across the two images under matching. This is normally accomplished by identifying those feature points with most similar local features represented by feature descriptors and measuring the feature-vector distance based on the nearest neighbor matching criterion. However, a large number of mismatches would be incurred especially when the two images under comparison have a large viewpoint variation with respect to each other or involve very different backgrounds. In this paper, a new mismatch removal method is proposed by utilizing the bipartite graph to first establish one-to-one coherent region pairs (CRPs), which are then used to verify whether each point-to-point matching pair is a correct match or not. The generation of CRPs is achieved by applying the Hungarian method to the bipartite graph, together with the use of the proposed region-to-region similarity measurement metric. Extensive experimental results have demonstrated that our proposed mismatch removal method is able to effectively remove a significant number of mismatched point-to-point correspondences.