Learning Cost Functions for Graph Matching

During the last decade, several approaches have been proposed to address detection and recognition problems, by using graphs to represent the content of images. Graph comparison is a key task in those approaches and usually is performed by means of graph matching techniques, which aim to find correspondences between elements of graphs. Graph matching algorithms are highly influenced by cost functions between nodes or edges. In this perspective, we propose an original approach to learn the matching cost functions between graphs' nodes. Our method is based on the combination of distance vectors associated with node signatures and an SVM classifier, which is used to learn discriminative node dissimilarities. Experimental results on different datasets compared to a learning-free method are promising.