This repository is about implementing the paper Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition Arxiv Preprint using the Pytorch Geometric package. A series of experiments will be conducted to test the various methods and to see which method gives the best result. The dataset used is NTU-60 RGBD.
- Experiment 1: The sequences are padded along the temporal dimension to a common length.
- Experiment 2: The data is padded along the feature dimension.
- Experiment 3: The data for 2 person is batched together along the temporal dimension such that both the graphs are treated separately for spatial and temporal convolution.
- Experiment original stgcn: The original code is trained from scratch to observe the training and validation plots.
- Experiment 4: The number of stgcn blocks are kept constant (10) and the hyperparameters are tuned.
- Experiment 5: The number of stgcn blocks are varied and the accuracies are observed.
- Experiment 6: The issue of isotropic kernel of GCN is identified.
| Date | Experiment | Epochs | Training Accuracy | Validation accuracy |
|---|---|---|---|---|
| 7/12/2020 | Exp 1 | 80 | 0.23 | 0.016 |
| 10/12/2020 | Exp 2 | 80 | 0.66 | 0.57 |
| 14/12/2020 | Exp 3 | 80 | 0.71 | 0.60 |
| 15/12/2020 | Exp 3b | 80 | 0.79 | 0.51 |
| 18/12/2020 | Exp original_stgcn | 80 | 0.9993 | 0.7751 |
Since Expt 4,5,6 have sub experiments go to those folders to view the table of accuracies.
- Python (3.6.8)
- Pytorch (1.6.0)
- Torch-Geometric (1.6.1)
- Torch-Cluster (1.5.7)
- Torch-Scatter (2.0.5)
- Torch-Sparse (0.6.7)
- Torch-Spline-conv (1.2.0)