Bootstrapping Joint Bayesian model for robust face verification

Abstract

Generative Bayesian models have exhibited good performance on the face verification problem, i.e., determining whether two faces are from the same person. As one of the most representative methods, the Joint Bayesian (JB) model represents two faces jointly by introducing some appropriate priors, providing better separability between different face classes. The EM-like learning algorithm of the JB model, however, are occasionally observed to have unsatisfactory converge property during the iterative training process. In this paper, we present a Bootstrapping Joint Bayesian (BJB) model which demonstrates good converging behavior. The BJB model explicitly addresses the classification difficulties of different classes by gradually re-weighting the training samples and driving the Bayesian models to pay more attentions to the hard training samples. Experiments on a new challenging benchmark demonstrate promising results of the proposed model, compared to the baseline Bayesian models.