Efficient Multi-scale Network with Learnable Discrete Wavelet Transform for Blind Motion Deblurring

News

2024.02.28 Our Paper is accepted by CVPR2024.

Easy Start

1.Environmental settings

cd MLWNet
pip install -r requirements.txt
python setup.py develop --no_cuda_ext

2.Preparing the dataset

Download the required dataset and place it under ‘./dataset' and convert it to lmdb format. In the future, we will also open a direct link to download the dataset in lmdb format.

3.Start training

Take RSBlur and 8 GPUs as an example:

nohup python -m torch.distributed.launch --nproc_per_node=8 --master_port=4321 basicsr/train.py --opt options/train/RSBlur/MLWNet-B.yml --launcher pytorch &>MLWNet-B-RS.out&

4.Evaluation

We will be open-sourcing our pre-trained models. For RealBlur and RSBlur, please use their official alignment codes respectively.

python eval.py --weights your_weights --dir your_dataset --device your_device

5.Other

Pretrained models

---

GOPRO

width_64 Google Driver link: https://drive.google.com/file/d/15SjPtkfQ0m6NuVss0rwvscatKwRZ2l4i/view?usp=drive_link

---

RealBlur-J

width32 Google Driver link: https://drive.google.com/file/d/1WYhczcWj9vPn-C3fCfLk5Tq0GkD3ylY3/view?usp=drive_link

width64 Google Driver link: https://drive.google.com/file/d/1lv3MCJgZUWgITUFNvgY8nU9YlpgE050W/view?usp=drive_link

---

RealBlur-R

width64 Google Driver link: https://drive.google.com/file/d/1hp4Qu2n_lOmc7LsSOHYjf4NhUzZtskGP/view?usp=drive_link