Large Margin Loss for Learning Facial Movements from Pseudo-Emotions

In this paper we propose a large margin based loss function that enables information transfer from an unsupervised domain to a supervised one. The proposed methodology is applied in the context of face expression analysis. Categorical expressions are easier to understand and mutually exclusive, yet annotation is difficult and arguable. In contrast, facial movements encoded as action units have gained wider acceptance. Our strategy assumes self labelling images in the wild with pseudo-emotions to better learn action units. The proposed method is tested in two challenging scenarios with expressions in the wild, showing improved performance with respect to the baseline.