Partial correspondence based on subgraph matching

Exploiting both appearance similarity and geometric consistency is popular in addressing the feature correspondence problem. However, when there exist outliers the performance generally deteriorates greatly. In this paper, we propose a novel partial correspondence method to tackle the problem with outliers. Specifically, a novel pairwise term together with a neighborhood system is proposed, which, together with the other two pairwise terms and a unary term, formulates the correspondence to be solved as a subgraph matching problem. The problem is then approximated by the recently proposed Graduated Non-Convexity and Graduated Concavity Procedure (GNCGCP). The proposed algorithm obtains a state-of-the-art accuracy in the existence of outliers while keeping O(N3) computational complexity and O(N2) storage complexity. Simulations on both the synthetic and real-world images witness the effectiveness of the proposed method.