A perceptual organization approach to contour estimation via composition, compression and pruning of contour hypotheses

The problem of estimating scene contours by means of perceptual organization is formulated in a probabilistic fashion. The goal of estimation is to compute a set of contour descriptors which approximate every scene contour with high probability. A hierarchy of contour descriptors designed for this purpose is proposed. Computation at each level of the hierarchy consists of three basic ingredient: hypothesis generation by means of grouping; hypothesis evaluation and pruning; and, where necessary, compression of equivalent multiple responses. A specific algorithm to enumerate, prune and *Research supported by MURI grant DAAH04-96-1-0445, Foundations of Performance Metrics for Object Recognition. compress contour-cycles in a graph is proposed. This algorithm can easily incorporate feedback loops to prune the search which exploit global information.