Bayesian Identity Clustering

Our goal is to establish how many different people are present in a set of N facial images, and determine the correspondence between people and images. Our approach is Bayesian: in the training phase, we learn a probabilistic generative model for face data. Individual identity is represented as a latent variable in this model, and is constrained to be identical when faces match. We use this model to calculate the likelihood for the whole dataset for each hypothesized clustering: using a process equivalent to Bayesian model selection, we marginalize over the unknown identity variables allowing us to compare models with differing numbers of people. For large datasets, it is not possible to exhaustively examine every possible clustering, and we introduce approximate algorithms to cope with this case. We demonstrate results both for frontal faces, and for face sets containing large pose variations. We present a detailed quantitative evaluation of the results for a standard dataset.