Awesome

AccFlow: Backward Accumulation for Long-Range Optical Flow

ICCV2023 | AccFlow: Backward Accumulation for Long-Range Optical Flow
Guangyang Wu, Xiaohong Liu, Kunming Luo, Xi Liu, Qingqing Zheng, Shuaicheng Liu, Xinyang Jiang, Guangtao Zhai, Wenyi Wang

TODO:

• Add inference and visualization codes.
• Add a demo video for better understanding.
• Add figures and brief introduction of this work.
• Provide google drive link for CVO dataset.
• Add warmstart mode for evaluation.
• Add evaluation using GMFlow models.

Requirements

conda env create -f environment.yml
conda activate accflow

Models

We provide pretrained models. The default path of the models for evaluation is:

├── checkpoints
    ├── acc+raft-things.pth
    ├── acc+gma-things.pth
    ├── acc+raft-cvo.pth
    ├── acc+gma-cvo.pth
    ├── raft-cvo.pth
    ├── gma-cvo.pth
    ├── raft-things.pth
    ├── gma-things.pth

Evaluation

Download checkpoints and put it in the root dir.

Download testing dataset CVO-test, put the files cvo-test.lmdb and cvo-test.lmdb-lock in the directory data/datasets.

To evaluate on the clean and final splits, use '-d' param to specify. To evaluate direct methods (e.g., RAFT, GMA), set '-acc' to 'direct'. To evaluate accumulation methods (i.e., accflow), set '-acc' to 'acc'.

python test_cvo.py -d clean -acc direct -ofe raft --ofe_ckpt checkpoints/raft-things.pth
python test_cvo.py -d clean -acc acc -ofe raft --acc_ckpt checkpoints/acc+raft-things.pth

More samples can be found in test_cvo.sh.

Training

The script will load the config according to the training stage. The trained model will be saved in a directory in logs and checkpoints. For example, the following script will load the config configs/***.yml.

# Fine-tune RAFT and GMA (pretrained on flyingthings) using CVO training set
python fine_tune.py -c configs/RAFT.yml
python fine_tune.py -c configs/GMA.yml

# Train AccFlow based on RAFT and GMA (pretrained on flyingthings) using CVO training set
python train_acc.py -c configs/AccGMA.yml
python train_acc.py -c configs/AccRAFT.yml

# Train AccFlow based on RAFT and GMA (fine-tuned with CVO-train)
python train_acc.py -c configs/AccGMA-CVO.yml
python train_acc.py -c configs/AccRAFT-CVO.yml

License

AccFLow is released under the MIT License

Citation

If you use any part of this code, please kindly cite
@article{wu2023accflow,
  title={AccFlow: Backward Accumulation for Long-Range Optical Flow},
  author={Guangyang Wu and Xiaohong Liu and Kunming Luo and Xi Liu and Qingqing Zheng and Shuaicheng Liu and Xinyang Jiang and Guangtao Zhai and Wenyi Wang},
  journal={arXiv preprint arXiv:2308.13133},
  year={2023}
}

Acknowledgement

In this project, we use parts of codes in:

• RAFT
• GMA