Universität Bielefeld - Sonderforschungsbereich 360

Generic Shape Indexing and Matching for 2-D Object Recognition

Sven Dickinson
Department of Computer Science and Center for Cognitive Science
Rutgers University

Tuesday, May 19th, 1998
16.15 h
Room: Hörsaal 5

We have been developing a theory for the generic representation of 2-D shape, in which structural descriptions are derived from the shocks (singularities) of a curve evolution process acting on a silhouette. We now apply the theory to the problems of shape matching and indexing. The shocks are organized into a directed, acyclic shock graph which, in turn, can be reduced to a unique rooted shock tree whose levels are ordered by significance. We first introduce a novel tree indexing algorithm based on an eigenvalue characterization of a shock tree's topological structure. This topological signature, in the form of a vector, is computed at each node of a query tree and used to "vote" for objects having similar local topological structure. From the resulting set of model shock tree candidates, this same topological signature is exploited in a shock tree matching (verification) algorithm that factors in the geometry of the shock tree's nodes. The algorithm computes the distance between two shock trees as well as a set of node correspondences. Using a diverse database of shapes, we demonstrate our system's performance under articulation, occlusion, and changes in viewpoint.

This talk represents joint work with Ali Shokoufandeh (Rutgers University), Kaleem Siddiqi (Yale University), and Steven W. Zucker (Yale University).

Anke Weinberger, 1998-03-02, 1998-04-02