A variational statistical framework for clustering human action videos

Abstract

In this paper, we present an unsupervised learning method, based on the finite Dirichlet mixture model and the bag-of-visual words representation, for categorizing human action videos. The proposed Bayesian model is learned through a principled variational framework. A variational form of the Deviance Information Criterion (DIC) is incorporated within the proposed statistical framework for evaluating the correctness of the model complexity (i.e. number of mixture components). The effectiveness of the proposed model is illustrated through empirical results.