Multimodal Facial Expression Recognition Based on Dempster-Shafer Theory Fusion Strategy

In this paper, we present our facial expression recognition framework for the Multimodal Emotion Recognition Challenge 2017. The task of the challenge is to recognize one of eight facial emotions in short video clips extracted from Chinese films, TV plays and talk shows. Considering the fact that each video is given only one video-level label, while each frame of the video is unlabeled, we improved the unsupervised auto-coder and used it for feature extraction. In addition, we have extracted other features, including traditional hand-craft features and deep learning features. Finally, the random forest algorithm was used to classify different emotional labels, and followed by a decision level fusion method based on Dempster-Shafer evidence theory. The macro average precision(MAP) of our best submission result achieves 59.68% on the testing set, which significantly outperforms the baseline of 30.63%.