Visual Saliency Maps Can Apply to Facial Expression Recognition

Human eyes concentrate different facial regions during distinct cognitive activities. We study utilising facial visual saliency maps to classify different facial expressions into different emotions. Our results show that our novel method of merely using facial saliency maps can achieve a descent accuracy of 65\%, much higher than the chance level of $1/7$. Furthermore, our approach is of semi-supervision, i.e., our facial saliency maps are generated from a general saliency prediction algorithm that is not explicitly designed for face images. We also discovered that the classification accuracies of each emotional class using saliency maps demonstrate a strong positive correlation with the accuracies produced by face images. Our work implies that humans may look at different facial areas in order to perceive different emotions.