Learning Organization Hierarchies of Large Model-Bases for Fast Recognition

Abstract

Abstract In attempting to realize a complete model-based vision system for CAD objects in range data, the two most critical issues are speed and accuracy in recognition. These factors become exceedingly important with increases in the size of the model library and the scene clutter. Past work in model-based vision attempted to design algorithms such that the recognition time would be insensitive to the model-base size or scene clutter [2, 4, 5, 6, 7, 8, 9, 10, 12, 13, 21]. Systems using point-based features to describe models (4, 6, 9, 12, 13] usually implement indexing ideas for object recognition. In the off-line model compilation process, indexes are calculated from a group of points, and these are used as addresses into a look-up table, where the identities of the model and the group of points are stored. Recognition entails a similar computation of addresses from groups of scene points, followed by voting for the models placed at these locations. This approach has become well-known under the moniker “geometric hashing.” Systems that use higher level primitives (like surfaces, volumes, etc.) traditionally represent both model and scene objects in the form of graphs (2, 7, 8, 10]. The nodes in these graphs represent the primitives, and the links represent the relationships among them. In recognition, the mapping function between the nodes of the scene graph and the nodes of each model graph is determined.