The Parallel EM Algorithm and its Applications in Computer Vision

* This work has been supported by the Spanish CICYT under grants TIC97-0897-C04-03 and TIC98-0559. Abstract In this paper we propose several applications of the EM algorithm –a well-known algorithm for parameter estimation in generic statistical problems– in the different levels of a computer vision system, and discuss its relatively straightforward and efficient implementation in parallel distributed memory environments, under the message passing paradigm. We show, at least, an algorithm for each level of a generic computer vision system (low, medium and high level vision) with the common basis of the EM algorithm applied in parameter estimation for mixture statistical models. We present an efficient implementation in parallel for all of them under a common scheme. Also, we show an evaluation of a medium-level vision EM application in a cluster of PCs. The results we have obtained in terms of both speedup and execution time are very good. Moreover, the paper gives us the guidelines for extending the method for related problems, and, in general, for many applications with mixture parameter estimation with EM as its main task.