HackMD
tags: GAS-GCN
Gated Action-Specific Graph Convolutional Networks for Skeleton-Based Action Recognition
About the paper
- read
- author: Wensong Chan ,Zhiqiang Tian, and Yang Wu
- published: 2020/6/21
- contributin
- ASGCN: combine structural and implicit edge
- Gated CNN: filter out useless temporal information
- Dataset:
- Experiments Details
- PyTorch: implement GAS-GCN
- SGD: optimization strategy
- 2 NVIDIA 2080TI GPU with memory of 11GB
- NTU-RGB + D training
- T=300
- epoch=50, base learning rate:
- 1~29: 0.1
- 30~40: 0.1 x 0.1
- 41~50: 0.1 x 0.1 x 0.1
- batch size=28
- Kinetics
- T=150
- epoch=65, base learning rate:
- 1~44: 0.1
- 45~55: 0.1 x 0.1
- 56~65: 0.1 x 0.1 x 0.1
- batch size=64
- 𝜇=0.5
GAS-GCN
- Gated Action-Specific Graph Convolutional Networks
- Why
- GCNs-based only focus on ajencent matrix (只有附近的joint，但是每個動作不一定只影響到一個關節附近)
- The connection between two joints is structural edge
- solution: action-specific graph convolutional module (ASGCM)
- generate the implicit edge
- decide the ratio of the combination of structural edge and implicit edges according to different action
- GAS-GCN
- Skeleton-based recognition depends on the contexts and long-range dependencies in temporal dimension
- 但不是所有資訊都會用到action recognition上
- solution: Gated mechanism
- control the information flowing in RNN
- operates the time dimension
