Pyskl Versions Save

New Algorithms

• AAGCN (TIP): https://arxiv.org/abs/1912.06971 [MODELZOO]
• MS-G3D (CVPR 2020 Oral): https://arxiv.org/abs/2003.14111 [MODELZOO]
• CTR-GCN (ICCV 2021): https://arxiv.org/abs/2107.12213 [MODELZOO]

New Datasets

• Kinetics-400 (2D skeleton)
• Diving48 (2D skeleton)

New Features

• Generating 2D skeleton annotations for a custom video dataset.
• Support spatial augmentations (rotate & scale).
• Support GCN inference in demo_skeleton

Bug Fix

• Fix bug in vis_heatmap.ipynb
• Fix links of ST-GCN++ weights

Known Issues

• Currently, demo_skeleton is not compatible with PoseC3D trained on K400.

Supported Algorithms

• PoseConv3D (CVPR2022 Oral): A 3D-CNN based approach for skeleton action recognition.
• ST-GCN (AAAI 2018): A GCN based approach for skeleton action recognition.
• ST-GCN++ (PYSKL): A variant of ST-GCN, which achieves outstanding recognition performance without any adaptive schemes.

Supported Datasets

• NTURGB+D & NTURGB+D 120 (all splits, both 2D & 3D skeletons).
• FineGYM (2D skeletons).
• UCF101 (2D skeletons).
• HMDB51 (2D skeletons).

Supported Features

• Training and testing of all supported algorithms.
• Visualization of both 2D and 3D skeleton annotations.