An Improved Action Recognition Network Based on Appearance and Relation

With the increase of people's security awareness, people need to distinguish human behavior in videos. In order to avoid manpower consumption, action recognition based on deep learning has been widely studied in recent year.3D convolution is used to improve computing efficiency for action recognition. In this paper, an improved action recognition network based on appearance and relation is proposed. The residual block in the ResNet3D network is redesigned into a new residual block with two branches of appearance and relation. In addition, the separable spatial/temporal convolution is used in the relation branch for more spatio-temporal information and less loss. The proposed architecture achieves comparable results to state-of-the-art methods on UCF101 .