Deep spatio-temporal feature fusion with compact bilinear pooling for multimodal emotion recognition

Abstract

Multimodal emotion recognition has attracted great interest recently and numerous methodologies have been successfully investigated. However, the task requires the effective fusion multimodal representations in audio and video domains, and existing approaches still perform poorly on such a challenging task. This paper proposes a novel framework for recognizing emotion from multiple sources including facial expression, pose, body movements, and voice. In this framework, we first introduce new deep spatio-temporal features by cascading 3-dimensional convolution neural networks (C3Ds) and deep belief networks (DBNs) to effectively model spatial and temporal information presented in video and audio for emotion recognition. We subsequently propose a new feature-level fusion approach based on a bilinear pooling theory to combine the visual and audio feature vectors. The proposed fusion strategy allows all elements of the component vectors to interact with each other in an effective way, resulting in expressively capturing the complex and intrinsic associations between the component modalities. Extensive experiments conducted on the eNTERFACE and FABO multimodal emotion databases demonstrate that our proposed system leads to improved multimodal emotion recognition performance and significantly outperforms recent state-of-the-art approaches.