FEATURE GROUPING IN A HIERARCHICAL PROBABILISTIC NETWORK

The goal of feature grouping is to provide efficient codings of the necessary information for much of the scene interpretation and object recognition applications in machine vision. This paper offers a sound theoretical background for feature grouping processes, using a Bayesian approach which makes explicit the world knowledge which is applied at any stage. We describe a framework which can integrate many different forms of grouping and different levels of information. We also report on a preliminary implementation within this framework to group parallel lines in a perspective image. In support of this we develop a mapping from the image uncertainty to the orientation uncertainty for the hypothesized groups.