Pairwise Ranking Network for Affect Recognition

In this work we study the problem of emotion recognition under the prism of preference learning. Affective datasets are typically annotated by assigning a single absolute label, i.e. a numerical value that describes the intensity of an emotional attribute, to each sample. Then, the majority of existing works on affect recognition employ sample-wise classification/regression methods to predict affe...