Two-Stream Adaptive Graph Convolutional Networks for Skeleton-Based Action Recognition

1, Topology of graph can be learned adaptively; (different samples and different layers have different topology)
2, Second-order information, ie, the lengths and directions of bones, are used for human action recognition;
3, Two stream network to model both the first-order and second-order information;

Adaptive graph convolutional layer

Adjacent matrix is divided into three parts: 2019 CVPR Two-Stream Adaptive Graph Convolution Networks for Human Action Recognition - 图1. 2019 CVPR Two-Stream Adaptive Graph Convolution Networks for Human Action Recognition - 图2 is normal adjacent matrix; 2019 CVPR Two-Stream Adaptive Graph Convolution Networks for Human Action Recognition - 图3 can be optimized by BP algorithm; 2019 CVPR Two-Stream Adaptive Graph Convolution Networks for Human Action Recognition - 图4 is calculated by estimating vertex similarity (as the figure shows below)
2019 CVPR Two-Stream Adaptive Graph Convolution Networks for Human Action Recognition - 图5 means matrix product, 2019 CVPR Two-Stream Adaptive Graph Convolution Networks for Human Action Recognition - 图6 means matrix addition.
image.png
image.png

Two-stream networks

image.png

Experiment results

image.png
image.png