Scalable data parallel implementations of object-recognition on Connection Machine CM-5

Object recognition involves identifying known objects in a given scene. It plays a key role in image understanding. Geometric hashing is a technique for model-based object recognition in occluded scenes. We present scalable data parallel algorithms for geometric hashing on a CM-5. Given a scene consisting of S feature points, the parallel algorithm for one probe of the recognition phase takes O/spl lsqb/(S/P) log S/spl rsqb/ time on a fat tree-based architecture. We perform implementations of the proposed algorithms on a CM-5, after a careful study of its computation and communication characteristics. Earlier parallel implementations of the geometric hashing algorithm were carried out on a CM-2 using O(Mn/sup 3/) processors, where M is the number of models in the database and n is the number of features in each model. In these implementations, the number of processors is independent of the size of the scene but depends on the size of the model database which is usually very large. The algorithms presented significantly improve on the number of processors employed, while at the same time achieving much superior time performance. Earlier implementations claim 700 to 1300 ms for one probe of the recognition phase, assuming 200 feature points in the scene on an 8K processor CM-2 Our implementations run on a P processor CM-5, such that 1/spl les/P/spl les/S. Our results show that a probe of the recognition phase for a scene consisting of 1024 feature points takes less than 10 ms on a a 256 processor CM-5. The implementations developed in this paper require number of processors independent of the size of the model database and are also scalable with the machine size.<
>