A New Implementation of k-MLE for Mixture Modeling of Wishart Distributions

We describe an original implementation of k-Maximum Likelihood Estimator (k-MLE)[1], a fast algorithm for learning finite statistical mixtures of exponential families. Our version converges to a local maximum of the complete likelihood while guaranteeing not to have empty clusters. To initialize k-MLE, we propose a careful and greedy strategy inspired by k-means++ which selects automatically cluster centers and their number. The paper gives all details for using k-MLE with mixtures of Wishart (WMMs). Finally, we propose to use the CauchySchwartz divergence as a comparison measure between two WMMs and give a general methodology for building a motion retrieval system.