Action Capsules: Human skeleton action recognition

Ali Farajzadeh Bavil, Hamed Damirchi, Hamid D. Taghirad

Abstract: Highlights•Action Capsules neural networks for skeleton-based human action recognition are proposed.•We explicitly model the latent correlation between human joints and actions.•We propose a method to dynamically attend to distinct action-related joints.•We achieve state-of-the-art accuracy on the N-UCLA benchmark dataset.•The most efficient network for action recognition considering its GFLOPs is presented.