Temporal Perceptive Network for Skeleton-Based Action Recognition

The major challenge for skeleton-based action recognition is to distinguish the difference between various actions. Traditional Recurrent Neural Network (RNN) structure may lead to unsatisfactory results due to the inefficiency in capturing local temporal features, especially for large-scale datasets. To address this issue, we propose a novel Temporal Perceptive Network (TPNet) to enable the robust feature learning for action recognition. We design a temporal convolutional subnetwork, which can be embedded between RNN layers, to enhance automatical feature extraction for local temporal dynamics. Experiments show that the proposed method achieves superior performance to other methods and generates new state-of-the-art results. The model won the first place in the ACCV Workshop Large-Scale 3D Human Activity Analysis Challenge in Depth Videos.