Spontaneous Emotion Recognition in Response to Videos

Abstract

In order to understand human emotions correctly taking into account only facial expressions, we are conducting the experiments on the spontaneous emotional facial videos of people watching musical video clips from DEAP open source dataset. We are reporting the comparative results of emotion recognition done in two ways: sequential extraction of spatial and temporal features done by CNN-RNN, simultaneous extraction of both types of features performed by 3D convolutions in our C3D networks architecture. In order to study the contribution of microexpressions to emotion recognition we are augmenting videos in two ways: reducing to 1 fps, thus losing a significant amount of temporal information, reducing to 10 fps, thus preserving most of the muscle movement information.