Transitive assignment kernels for structural classification

Abstract

Kernel methods provide a convenient way to apply a wide range of learning techniques to complex and structured data by shifting the representational problem from one of finding an embedding of the data to that of defining a positive semi-definite kernel. One problem with the most widely used kernels is that they neglect the locational information within the structures, resulting in less discrimination. Correspondence-based kernels, on the other hand, are in general more discriminating, at the cost of sacrificing positive-definiteness due to their inability to guarantee transitivity of the correspondences between multiple graphs. In this paper we adopt a general framework for the projection of (relaxed) correspondences onto the space of transitive correspondences, thus transforming any given matching algorithm onto a transitive multi-graph matching approach. The resulting transitive correspondences can then be used to provide a kernel that both maintains locational information and is guaranteed to be positive-definite. Experimental evaluation validates the effectiveness of the kernel for several structural classification tasks.