The Fusion Knowledge of Face, Body and Context for Emotion Recognition

Emotion recognition is a hot research direction in computer vision and is also an important way to perceive human behaviors. Traditional methodology to handle emotion recognition only captures the information from human face. In this paper, we introduce a novel framework to improve the performance of emotion recognition, from not only the clues in the face, but the body posture and environmental context. Meanwhile, we design a lighter network structure to reduce the computational load and a hierarchical sampling algorithm to improve the unbalance in training data. Experimental results show that our framework outperforms the method that only captures the emotion information from the face or body posture.