Two Stage Emotion Recognition using Frame-level and Video-level Features

This paper compares a seven class classifier with a two stage classification for categorical emotion recognition. We use the Multimodal Emotion Recognition (MER) Dataset of the FG2020 competition which apart from video contains skeleton data collected with a Microsoft Kinect. We compare the performance of different unimodal features as well as various combinations of multimodal features. We also compare frame-level features with video-level features. We achieved 50parcent accuracy using multimodal video-level features and two stage classification on one hand, 49 % accuracy is achieved with the seven class classifier on the other hand.