Audio-Video Based Emotion Recognition Using Minimum Cost Flow Algorithm

In this paper, we present an efficient system for Audio-Video based emotion recognition with a limited dataset in the wild. To recognize the emotion, we use both acoustic and facial features mode together. To deal with Audio-Video data, we utilize both temporal and non-temporal information. To solve the problem of a small amount of dataset, we experiment from conventional methods to the deep learning-based methods. Finally, to address the unbalanced and skewed distribution problems, we apply a graph theory called Minimum Cost Flow Algorithm. By those approaches, our methods perform 61.56% on the test set in Audio-Video Emotion Recognition sub-challenge of 2019 Emotion Recognition in the Wild (Emotiw) Challenge and rank 5th among several teams.