Unsupervised Parameterisation of Gaussian Mixture Models

In this paper we explain a new practical methodology to fully parameterise Gaussian Mixture Models (GMM) to describe data set distributions. Our approach analyses hierarchically a data set distribution to be modeled, determining unsupervisedly an appropriate number of components of the GMM, and their corresponding parameterisation. Results are provided that show the improvement of our method with regard to an implementation of the traditional approach usually applied to solve this problem. The method is also tested in the unsupervised generation of shape models for visual tracking applications.