A unified formulation of invariant point pattern matching

Abstract

We present a unified framework for modeling and solving invariant point pattern matching problems. Invariant features are encoded as potentials in a probabilistic graphical model. By using a specific kind of graph topology, different types of invariant matching models can be implemented via tree-width selection. Models with tree-widths 1, 2, 3 and 4 implement translation, similarity, affine and protective invariant point matching, respectively. The optimal match is then found by exploiting the Markov structure of the graph through the generalized distributive law in a dynamic programming setting. In the absence of noise in the point coordinates, the solutions found are optimal. Our early experiments suggest the approach is robust to outliers and moderate noise. Show more