论文阅读笔记:“Coherence Constrained Graph LSTM for Group Activity Recognition”

Coherence Constrained Graph LSTM for Group Activity Recognition (TPAMI 2022)
主要思想:寻找群体体行为中的相关行为
STCC和GCC约束
- STCC:如果一个个体行为在时间域上相互关联,并且在空间域上与其他个体行为具有一致性,我们称该种行为为相关行为
- GCC:如果一个特定的行为与整体群体行为一致性更高,那么它对群体行为的贡献也越高

网络框架

核心模块
spatial interactions among persons are dependent on the temporal motions of individuals
CCG-LSTM
时空场景关联推理




temporal confidence gate


- 如果相差越小,值越接近1
- 加入这个temporal confidence gate后的新公式:

spatial context confidence gate
捕获空间域的个体的“relative”运动



- 如果相差越小,值越接近1
- 加入这个spatial confidence gate后的新公式:

AGG-LSTM
全局场景关联推理

- Post title:论文阅读笔记:“Coherence Constrained Graph LSTM for Group Activity Recognition”
- Post author:sixwalter
- Create time:2023-04-17 00:00:00
- Post link:https://coelien.github.io/2023/04/17/paper-reading/paper_reading_063/
- Copyright Notice:All articles in this blog are licensed under BY-NC-SA unless stating additionally.
Comments