CycleACR: Cycle Modeling of Actor-Context Relations for Video Action Detection
Lei Chen, Zhan Tong, Yibing Song, Gangshan Wu, Limin Wang
2025.09.11The code and weights for VideoMAE-ViT-L backbone are available!2025.08.11Code and pre-trained models are available now!2025.07.30Our CycleACR is accepted by T-PAMI 2025! 🎉
- For SlowFast backbone: Please follow the installation instructions in INSTALL.md.
- For VideoMAE backbone: Please refer to VideoMAE-Action-Detection/INSTALL.md for installation instructions.
Both backbones share the same data preparation process. Please follow the instructions in VideoMAE-Action-Detection/DATASET.md to prepare the AVA dataset.
| method | config | backbone | pre-train | AVA mAP | model |
|---|---|---|---|---|---|
| CycleACR | cfg | SlowFast-R101-8x8 | K700 | 34.0 | link |
| CycleACR | script | VideoMAE-ViT-L | K700 | 40.6 | link |
Note: The fine-tuning instruction for VideoMAE-ViT-L backbone is in FINETUNE.md.
python -m torch.distributed.launch --nproc_per_node=8 train_net.py --config-file "config_files/config_file.yaml" --transfer --no-head --use-tfboard --skip-final-test
python -m torch.distributed.launch --nproc_per_node=8 test_net.py --config-file "config_files/config_file.yaml" MODEL.WEIGHT "path/to/model/weight"
This project is built upon AlphaAction and maskrcnn-benchmark. Thanks to the contributors of these great codebases. We also thankfully acknowledge the computing resource support of Tencent Corporation for this project.
If you find this project useful, please feel free to leave a star and cite our paper:
@article{chen2025cycleacr,
title={Cycleacr: Cycle modeling of actor-context relations for video action detection},
author={Chen, Lei and Tong, Zhan and Song, Yibing and Wu, Gangshan and Wang, Limin},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
year={2025},
publisher={IEEE}
}
@article{chen2023cycleacr,
title={CycleACR: Cycle Modeling of Actor-Context Relations for Video Action Detection},
author={Chen, Lei and Tong, Zhan and Song, Yibing and Wu, Gangshan and Wang, Limin},
journal={arXiv preprint arXiv:2303.16118},
year={2023}
}