Learning structural models in multiple projection spaces

Abstract

We present an Expectation-Maximization learning algorithm (E.M.) for estimating the parameters of partially-constrained Bayesian trees. The Bayesian trees considered here consist of an unconstrained subtree and a set of constrained subtrees. In this tree structure, constraints are imposed on some of the parameters of the parametrized conditional distributions, such that all conditional distributions within the same subtree share the same constraint. We propose a learning method that uses the unconstrained subtree to guide the process of discovering a set of relevant constrained tree substructures. Substructure discovery and constraint enforcement are simultaneously accomplished using an E.M. algorithm. We show how our tree substructure discovery method can be applied to the problem of learning representative pose models from a set of unsegmented video sequences. Our experiments demonstrate the potential of the proposed method for human motion classification. © 2009 Springer Berlin Heidelberg.