Meta Transferring for Deblurring (BMVC 2022)

Pytorch implementation of "Meta Transferring for Deblurring" (BMVC 2022).
The implementation use part of the code from MPRNet, MTRNN, Restormer, and CDVD-TSP.

Environment

• Nvidia A5000 GPU
• cuda 11.3
• pytorch == 1.10.1+cu113
• numpy == 1.22.0

Installation

pip3 install -r requirements.txt

Meta training

1. Download "GOPRO" dataset or any dataset that you want to use.
2. Move pre-trained weights into ./weights.
3. Modify the arguments in meta training script and run the script.

sh meta_train.sh

Meta testing

1. Download "DVD", "REDS",or "RealBlur" dataset or any dataset that you want to test.
2. Modify the arguments in meta testing script and run the script.

sh meta_test_script.sh

Important args

• save_dir: The saved path of current experiment directory
• dataset: The dataset name (GOPRO/DVD/REDS/RealBlur)
• dataset_dir: The path to dataset
• deblur(reblur)_lr: Learning rate for deblurring(reblurring) model
• deblur_model: The deblurring model name (mprnet/mtrnn/restormer/cdvd_tsp)
• gan: Using adversarial loss during inner update or not
• cycle_update: Using cycle consistency loss during inner update or not
• n_updates: # of support patches during inner update
• deblur(reblur/gan)_model_path: The path of pre-trained model weight

Train/Test reblurring model

We train our reblurring model for 1000 epochs.

# Training
sh train.sh

# Testing
sh test.sh

Model/Dataset extension

• Model

  • Modify ./model/setting.yaml
  • Modify ./model/basicModule.py
  • Modify ./loss/loss.py

• Dataset

  • Create ./dataset/meta_(modelName).py
  • Copy ./dataset/meta_gopro.py to ./dataset/meta_(modelName).py
  • Modify dataset hierachy in the file

Citation

@inproceedings{liu2022meta,
  title={Meta Transferring for Deblurring},
  author={Liu, Po-Sheng and Tsai, Fu-Jen and Peng, Yan-Tsung and Tsai, Chung-Chi and Lin, Chia-Wen and Lin, Yen-Yu},
  booktitle={BMVC},
  year={2022}
}