Massively parallel model matching: geometric hashing on the Connection Machine

Abstract

The parallelizability of geometric hashing is explored, and two algorithms are presented. Geometric hashing uses the collection of models in a preprocessing phase (executed off line) to build a hash table data structure. The data structure encodes the model information in a highly redundant, multiple-viewpoint way. During the recognition phase, when presented with a scene and extracted features, the hash table data structure indexes geometric properties of the scene features to candidate models. The first uses: parallel hypercube techniques to route information through a series of maps and building-block parallel algorithms. The second algorithm uses the Connection Machine's large memory resources and achieves parallelism through broadcast facilities from the front end. The discussion is confined to the problem of recognizing dot patterson embedded in a scene after they have undergone translation, rotation, and scale changes.