任务
action recognition
motivation
Both recurrent and convolutional operations are neighborhood-based local operations either in space or time; hence local-range information is repeatedly extracted and propagated to capture long-range dependencies. Many works have designed networks with hierarchical structure to obtain longer range and deeper semantic information but the problem still persists
if there are back and forth semantic dependencies.
基本流程
先用全连接神经网络或者卷积神经网络把骨架节点的raw position and velocity sequences变成encoded features，再把encoded features放进自注意网络(self-atttion network, SAN)中处理。SAN有很多个版本，如下图所示
