Unidirectional Video Denoising by Mimicking Backward Recurrent Modules with Look-ahead Forward Ones

This source code for our paper "Unidirectional Video Denoising by Mimicking Backward Recurrent Modules with Look-ahead Forward Ones" (ECCV 2022)

Usage

Dependencies

You can create a conda environment with all the dependencies by running

conda env create -f requirements.yaml -n <env_name>

Datasets

For synthetic gaussian noise, DAVIS-2017-trainval-480p dataset is used for training, DAVIS-2017-test-dev-480p and Set8 are used for testing. For real world raw noise, CRVD dataset is used for training and testing.

Testing

Download pretrained models from Google Drive or Baidu Netdisk. We also provide denoised results (tractor from DAVIS-2017-test-dev-480p) for visual comparison.

For synthetic gaussian noise,

cd test_models
python sRGB_test.py \
    --model_file <path to model file> \
    --test_path <path to test dataset>

For real world raw noise,

cd test_models
python CRVD_test.py \
    --model_file <path to model file> \
    --test_path <path to CRVD dataset>

Training

For synthetic gaussian noise,

cd train_models
python sRGB_train.py \
    --trainset_dir <path to train dataset> \
    --valset_dir <path to validation set> \
    --log_dir <path to log dir>

For real world raw noise,

cd train_models
python CRVD_train.py \
    --CRVD_dir <path to CRVD dataset> \
    --log_dir <path to log dir>

For distributed training of synthetic gaussian noise,

cd train_models
python -m torch.distributed.launch --nproc_per_node=4 sRGB_train_distributed.py \
    --trainset_dir <path to train dataset> \
    --valset_dir <path to validation set> \
    --log_dir <path to log dir>

Citation

If you find our work useful in your research or publication, please cite:

@article{li2022unidirectional,
  title={Unidirectional Video Denoising by Mimicking Backward Recurrent Modules with Look-ahead Forward Ones},
  author={Li, Junyi and Wu, Xiaohe and Niu, Zhenxing and Zuo, Wangmeng},
  booktitle={ECCV},
  year={2022}
}