Separable 3D residual attention network for human action recognition

As an important research issue in computer vision, human action recognition has been regarded as a crucial mean of communication and interaction between humans and computers. To help computers automatically recognize human behaviors and accurately understand human intentions, this paper proposes a separable three-dimensional residual attention network (defined as Sep-3D RAN), which is a lightweight network and can extract the informative spatial-temporal representations for the applications of video-based human computer interaction. Specifically, Sep-3D RAN is constructed via stacking multiple separable three-dimensional residual attention blocks, in which each standard three-dimensional convolution is approximated as a cascaded two-dimensional spatial convolution and a one-dimensional temporal convolution, and then a dual attention mechanism is built by embedding a channel attention sub-module and a spatial attention sub-module sequentially in each residual block, thereby acquiring more discriminative features to improve the model guidance capability. Furthermore, a multi-stage training strategy is used for Sep-3D RAN training, which can relieve the over-fitting effectively. Finally, experimental results demonstrate that the performance of Sep-3D RAN can surpass the existing state-of-the-art methods.