PyTorch implementation of "Modeling the Relative Visual Tempo for Self-supervised Skeleton-based Action Recognition", ICCV 2023 [PDF]
# generate raw database for NTU-RGB+D
python tools/ntu_gendata.py --data_path <path to nturgbd+d_skeletons>
python feeder/preprocess_ntu.py
## Unsupervised Pretraining
- Example for unsupervised pretraining of RVTCLR. You can change .yaml files in config/ntu60/pretext folder.
python main.py pretrain_skeletonclr --config config/ntu60/pretext/pretext_skeletonclr.yaml
python main.py pretrain_skeletonclr --config config/ntu60/pretext/pretext_skeletonclr_bone.yaml
python main.py pretrain_skeletonclr --config config/ntu60/pretext/pretext_skeletonclr_motion.yaml
## Linear Evaluation
- Example for linear evaluation of RVTCLR. You can change .yaml files in config/ntu60/linear_eval folder.
python main.py linear_evaluation --config config/ntu60/linear_eval/linear_eval_skeleton.yaml
python main.py linear_evaluation --config config/ntu60/linear_eval/linear_eval_skeleton_bone.yaml
python main.py linear_evaluation --config config/ntu60/linear_eval/linear_eval_skeleton_motion.yaml
## Acknowledgement
This repo is based on
- [CrosSCLR](https://github.com/LinguoLi/CrosSCLR)
- [AimCLR](https://github.com/Levigty/AimCLR)
- [AS-CAL](https://github.com/LZU-SIAT/AS-CAL)
## Citation
Please cite this work if you find it useful:
@InProceedings{Zhu_2023_ICCV, author = {Zhu, Yisheng and Han, Hu and Yu, Zhengtao and Liu, Guangcan}, title = {Modeling the Relative Visual Tempo for Self-supervised Skeleton-based Action Recognition}, booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)}, month = {October}, year = {2023}, pages = {13913-13922} }