In this paper we propose a method for recognizing facial expressions using information from a pair of domains: one has labelled data and one with unlabelled data. As the two domains may differ in distribution, we depart from the traditional semi--supervised framework towards a transfer learning approach. In our method, which we call Annealed Label Transfer, the deep learner explores and predicts labels on the unsupervised part, yet, in order to prevent too much confidence in its predictions (as domains are not identical), the global error is regularized with a randomization input via an annealing process. The method's evaluation is carried out on a set of four scenarios. The first two are standard benchmarks with expression faces in the wild, while the latter two have been little attempted before: face expression recognition in children and the study of the separability of anxiety-originated expressions in the wild. In all cases we show the superiority of the proposed method with respect to the strong baselines.
Supplementary material (ZIP)