Ghost Graph Convolutional Network for Skeleton-based Action Recognition

Skeleton-based action recognition has attracted great attention in human action recognition. Existing methods for skeleton-based action recognition improve performance by designing deeper networks without considering the efficiency of the model. In this paper, we propose a simple and effective lightweight graph convolutional network for skeleton-based action recognition. Our model is composed of a lightweight temporal convolutional network and spatial graph convolutional network using depthwise convolution. In addition, we propose a novel graph convolution that can take the multi-scale relationship of joints with low computational complexities. On the NTU RGB+D dataset, our proposed model achieves comparable or higher performance with much fewer parameters compared with baseline method.