Distance-GCN for Action Recognition

Haetsal Lee, Junghyun Cho, Ig-Jae Kim, Unsang Park

Abstract: Many skeleton-based action recognition models have been introduced with the application of graph convolutional networks (GCNs). Most of the models suggested new ways to aggregate adjacent joints information. In this paper, we propose a novel way to define the adjacency matrix from the geometrical distance between joints. By combining this method with the formerly used adjacency matrix, we can increase the performances of graph convolution layers with slightly increased computational complexity. Experiments on two large-scale datasets, NTU-60 and Skeletics-152, demonstrate that our model provides competitive performance.