Skeleton-based action recognition by part-aware graph convolutional networks

This paper proposes an improved graph convolutional networks to deal with the skeleton-based action recognition. Inspired by splitting skeleton into several parts to feed deep networks, the part-aware convolutions is designed to replace common convolutions which is performed on all the neighboring joints. For scale invariance on multi-scale data, an Inception-like structure is introduced, which can concatenate feature maps from different convolution kernels. In contrast to methods based on LSTMs, the model presented is capable of extracting both temporal and spatial features from input data. Due to full use of spatial structure, the performance is enhanced greatly on various datasets. To evaluate the model, experiments were conducted on three benchmark skeleton-based datasets, including Berkeley MHAD, SBU Kinect Interaction, and NTU RGB-D datasets. The effectiveness and robustness of the model are demonstrated by comparing the experimental results of the proposed model with the state-of-the-art results. In addition, feature maps from different layers of a trained model are explored and the explanation of the part-aware convolutions is also provided.