Domain Adaption for Facial Expression Recognition

Abstract

Facial expression recognition (FER) is a task that recognizes human emotions from their facial expressions. Owing to the lack of large datasets, a FER system is difficult to design, especially for real world environment. In this paper, we propose a new dataset augmentation method for FER and the corresponding training strategy by using similarity preserving generative adversarial network (SPGAN). By borrowing the idea from person re-ID field, we consider dataset augmentation as a domain adaptation task. The SPGAN is first trained on a lab condition dataset and a real world condition dataset to generate domain adapted images, and then CNN models are subsequently trained on the domain adapted images. We test our models on the RAF-DB and SFEW 2.0 datasets to show the improvement when compared it to our baseline. We also report our competitive accuracy when compared it with other state of the art works, which shows promissing results.