Exploiting sparsity and co-occurrence structure for action unit recognition

We present a novel Bayesian framework for facial action unit recognition. The first key observation behind this work is sparsity: out of possible 45 (and more) facial action units, only very few are active at any moment. The second is the strong statistical co-occurrence structure: most facial expressions are made by common combinations of facial action units, so knowing the presence of one can act as a strong prior for inferring the presence of others. We developed a novel Bayesian graphical model that encodes these two natural aspects of facial action units via compressed sensing and group-wise sparsity inducing priors. One crucial aspect of our approach is the allowance of overlapping group structures, which proves useful in dealing with action units that occur frequently across multiple groups. We derive an efficient inference scheme and show how such sparsity and co-occurrence can be automatically learned from data. Experiments on three standard benchmark datasets show superiority over the state-of-the-art.