A Hypergraph Matching Framework for Refining Multi-source Feature Correspondences

Abstract

In this paper, we develop a hypergraph matching framework which enables feature correspondence refinement for multi-source images. For images obtained from different sources (e.g., RGB images and infrared images), we first extract feature points by using one feature extraction scheme. We then establish feature point correspondences in terms of feature similarities. In this scenario, mismatches tend to occur because the feature extraction scheme may exhibit certain ambiguity in characterizing feature similarities for multi-source images. To eliminate this ineffectiveness, we establish an association hypergraph based on the feature point correspondences, where one vertex represents a feature point pair resulted from the feature matching and one hyperedge reflects the higher-order structural similarity among feature point tuples. We then reject the mismatches by identifying outlier vertices of the hypergraph through higher order clustering. Our method is invariant to scale variation of objects because of its capability for characterizing higher order structure. Furthermore, our method is computationally more efficient than existing hypergraph matching methods because the feature matching heavily reduces the enumeration of possible point tuples for establishing hypergraph models. Experimental results show the effectiveness of our method for refining feature matching.