Expectation propagation learning of finite Beta-Liouville mixtures for spatio-temporal object recognition

摘要

In this paper, we develop an efficient approach for the learning of finite Beta-Liouville mixture models. Unlike existing approaches, our is based on expectation propagation for parameters estimation and can select automatically the appropriate number of mixture components. We provide a coherent and unified learning framework to learn the complexity of the deployed mixture models and all the involved model parameters. We illustrate the performance of our learning algorithm with artificial data and a real application namely spatio-temporal objects (or dynamic events) recognition which has significant potential to be used in interactive systems or robotics. In particular, we highlight three of the most common spatio-temporal objects which involving facial expression, human activities and hand gesture. Our experiments results show the merits of the proposed approach.