Generalizing Integer Projected Graph Matching Algorithm for Outlier Problem

Abstract

Graph matching plays an important role in computer vision and pattern recognition. Recent graph matching algorithms usually formulate graph matching by a discrete optimization problem, and have designed various types of optimization techniques to find a local optimum in reasonable time. Among them some algorithms utilizing the graduated projection to the discrete domain exhibit superior performance, but these algorithms are limited to specific applications. From the outlier perspective, they are applicable to subgraph matching in which outliers exist in at most one graph. However, in real tasks there are usually outliers in both graphs. Previously we have proposed a method directly targeting at finding the most similar subgraphs in two weighted graphs. In this paper we show that the idea can be generalized to other algorithms, and the IPFP is chosen as a representative algorithm for generalization. Experiments witness the effectiveness of the generalization.