A Bayesian Approach to Learning Single View Generalization in 3D Object Recognition

Three dimensional vision relies on the ability to generalize from known views of an object to novel views. Of particular interest is the ability of human observers to generalize from a single view of a previously unseen object to novel views. This paper describes a method for achieving single view generalization by modeling conditional densities of the form P (S|B, B′) or P (B′|B, S) and applying them in a Bayesian decision theoretic framework, where B and B′ are two views and S is a boolean variable indicating whether the two views are of the same object or of different objects. Results on a standard set of test problems are given that demonstrate that such statistical models achieve considerably better single-view generalization ability than generalization based on 2D similarity alone. The approach can be used with many commonly used learning methods; in this paper, multi-layer perceptrons (MLP) and empirical distributions are used. Furthermore, the approach demonstrates object category-specific phenomena, similar to those observed psychophysically. Extensions of the approach to recognition from model bases and multiple example views are described, and interactions among multiple example views are explained as Bayesian combination of evidence among partially independent sources of evidence. The results presented in the paper suggest that such Bayesian models are a parsimonious and general approach to generalization in 3D vision.