A Probabilistic Model for Correspondence Problems Using Random Walks with Restart

Abstract

In this paper, we propose an efficient method for finding consistent correspondences between two sets of features. Our matching algorithm augments the discriminative power of each correspondence with the spatial consistency directly estimated from a graph that captures the interactions of all correspondences by using Random Walks with Restart (RWR), one of the well-established graph mining techniques. The $\it{steady}$-$\it{state}$ probabilities of RWR provide the global relationship between two correspondences by the local affinity propagation. Since the correct correspondences are likely to establish global interactions among them and thus form a strongly consistent group, our algorithm efficiently produces the confidence of each correspondence as the likelihood of correct matching. We recover correct matches by imposing a sequential method with mapping constraints in a simple way. The experimental evaluations show that our method is qualitatively and quantitatively robust to outliers, and accurate in terms of matching rate in various matching frameworks.