Game-theoretic hypergraph matching with density enhancement

Abstract

Feature matching plays a fundamental role in computer vision and pattern recognition. As straightforward comparison of feature descriptors is not enough to provide reliable matching results in many situations, graph matching makes use of the pairwise relationship between features to improve matching accuracy. Hypergraph matching further employs the relationship among multiple features to provide more invariance between feature correspondences. Existing hypergraph matching algorithms usually solve an assignment problem, where outliers may result in a large number of false matches. In this paper we cast the hypergraph matching problem as a non-cooperative multi-player game, and obtain the matches by extracting the evolutionary stable strategies. Our algorithm exerts a strong constraint on the consistency of obtained matches, and false matches are excluded effectively. In order to increase the number of matches without increasing the computation load evidently, we present a density enhancement method to improve the matching results. We further propose two methods to enforce the one-to-one constraint, thereby removing false matches and maintaining a high matching accuracy. Experiments with both synthetic and real datasets validate the effectiveness of our algorithm.