Geometric Hashing: Rehashing for Bayesian Voting I. Blayvas, R. Goldenberg, M. Lifshits, E. Rivlin, and M. Rudzsky Computer Science Department, Technion, Haifa, Israel Abstract. Geometric hashing is a model-based recognition technique based on matching of transformation-invariant object representations stored in a hash table. Today it is widely used as an object recognition method in numerous applications. In the last decade a number of en- hancements have been suggested to the basic method improving its per- formance and reliability. Here we consider two of them. One is rehashing, dealing with the problem of non-uniform occupancy of hash bins, and the other is Bayesian approach improving recognition rate in presence of noise. The latter uses an altered matching scheme, where the search is performed over an error dependent voting region around the query. In this paper we propose a scheme that takes the best from both worlds, yielding a hash table with a uniform size of voting regions. This allows to improve the geometric hashing computational performance by minimiz- ing the hash table size and the number of bins accessed, while maintain- ing optimal recognition rate. Alternatively, the proposed method can be used in classical single bin voting to improve recognition rate. Technion - Computer Science Department - Technical Report CIS-2003-07 - 2003