Action unit intensity estimation using hierarchical partial least squares

Estimation of action unit (AU) intensities is considered a challenging problem. AUs exhibit high variations among the subjects due to the differences in facial plasticity and morphology. In this paper, we propose a novel framework to model the individual AUs using a hierarchical regression model. Our approach can be seen as a combination of locally linear Partial Least Squares (PLS) models where each one of them learns the relation between visual features and the AU intensity labels at different levels of details. It automatically adapts to the non-linearity in the source domain by adjusting the learned hierarchical structure. We evaluate our approach on the benchmark of the Bosphorus dataset and show that the proposed approach outperforms both the 2D state-of-the-art and the plain PLS baseline models. The generalization to other datasets is evaluated on the extended Cohn-Kanade dataset (CK+), where our hierarchical model outperforms linear and Gaussian kernel PLS.