Online variational finite Dirichlet mixture model and its applications

Due to the increasing availability of digital data (e.g. image, text, video), online learning techniques have become much more desirable nowadays. This paper introduces an online algorithm for Dirichlet mixture models learning. By adopting the variational inference framework in an online manner, all the involved parameters and the model complexity of the Dirichlet mixture model can be estimated simultaneously in a closed form. Moreover, the problem of overfitting is prevented. The proposed algorithm is applied on two challenging real-world applications namely online object class recognition and online face tracking.