On the Foundations of Probabilistic Relaxationwith Product Support

Abstract

Traditional probabilistic relaxation, as proposed by Rosenfeld, Hummel and Zucker, uses a support function which is a double sum over neighboring nodes and labels. Recently, Pelillo has shown the relevance of the Baum-Eagon theorem to the traditional formulation. Traditional probabilistic relaxation is now well understood in an optimization framework. Kittler and Hancock have suggested a form of probabilistic relaxation with product support, based on an evidence combining formula. In this paper we present a formal basis for Kittler and Hancocks probabilistic relaxation. We show that it too has close links with the Baum-Eagon theorem, and may be understood in an optimization framework. We provide some proofs to show that a stable stationary point must be a local maximum of an objective function. We present a new form of probabilistic relaxation that can be used as an approximate maximizer of the global labeling with maximum posterior probability.