Spectral Correspondence for Deformed Point-Set Matching

This paper describes a modal method for point-set tracking in motion sequences. The framework for our study is the recently reported dual-step EM algorithm of Cross and Hancock [3]. This provides a statistical framework in which the structural arrangement of the point-sets provides constraints on the pattern of correspondences used to estimate alignment parameters. In this paper our representation of point-set structure is based on the point-adjacency matrix. Using ideas from spectral graph-theory, we show how the eigen-vectors of the point-adjacency matrix can be used to compute point correspondence probabilities. We show that the resulting correspondence matching algorithm can be used to track deforming point-sets detected in motion sequences.