Optimization of the Intensity Aware Loss for Dynamic Facial Expression Recognition

Dynamic facial expression recognition (DFER) is a challenging topic in computer vision. One of the critical problems in DFER is the variation in intensity in the different scenes. Different intensities lead to different traits and in the low intensity scenarios most emotions will be misclassified as neutral expression. The intensity aware loss has been introduced as a method to put more focus on the low intensity samples and better capture the traits of low intensity cases to acquire better discriminative abilities. In this work, an alternative way to define the intensity aware loss is proposed. With the use of the logarithmic Euclidian distance, the cross-entropy loss, and the exponential decrease model, and model combination, the discriminative ability in the low-intensity case can be further improved. Moreover, several ways to further optimize the model was proposed. Experiments show that, with these proposed techniques, even better performance of DFER can be achieved.