README.md

:circus_tent: Dataset Preparation

• Download the training GoPro dataset for constructing the synthetic dataset, run

python download_data.py --data train

• Generate image patches from full-resolution training images, run

python generate_patches_gopro.py 

• Download the training/testing RealBlur_J dataset for constructing the real-world dataset, refer to the download link.

• (Optional) Download the large-scale clean images for the unpaired conditions in diffusion, such as COCO or ImageNet datasets.

• Wrap the paths of the above datasets in .flist files for unified data loading, run

cd Noise-DA
python path_wrapper.py --folder_path 'dataset/deblur/train/GoPro/input_crops' --flist_file './flist_name/deblur/gopro_train_patch_input.flist'
python path_wrapper.py --folder_path 'dataset/deblur/train/GoPro/target_crops' --flist_file './flist_name/deblur/gopro_train_patch_gt.flist'
(Optional) python path_wrapper.py --folder_path 'dataset/COCO' --flist_file './flist_name/coco.flist'

• Split the training and testing sets from the RealBlur_J dataset using the predefined paths in ./datasets/split_names/, run

cd Deblurring
python data_split.py --source_file_path './datasets/split_names/RealBlur_J_train_list.txt' --output_file_path1 './flist_name/deblur/RealBlur_J_train_gt.flist' --output_file_path2 './flist_name/deblur/RealBlur_J_train_input.flist' --base_path './dataset/deblur/RealBlur_J/'
python data_split.py --source_file_path './datasets/split_names/RealBlur_J_test_list.txt' --output_file_path1 './flist_name/deblur/RealBlur_J_test_gt.flist' --output_file_path2 './flist_name/deblur/RealBlur_J_test_input.flist' --base_path './dataset/deblur/RealBlur_J/'

:dolphin: Training

• To train the image restoration network with the proposed diffusion loss, customize the training/validation paths (wrapped .flist files) of datasets in ./configs/options_train.json (from 26th to 28th lines and from 39th to 40th lines), and run

cd Noise-DA
sh train.sh Deblurring/configs/options_train.json

• (Optional) To train the image restoration network with the unpaired condition extension, customize the "ref_root" path of the large-scale clean images (e.g., ./flist_name/coco.flist, the 29th line) and set "diff_flag" to 2 (the 6th line) in ./configs/options_train.json, run

sh train.sh Deblurring/configs/options_train.json

Note: The above training scripts use 8 GPUs by default. To use any other number of GPUs, modify the values of -gpu in NoiseDA/train.sh.

:whale: Resume Training from Interruption

• Uncommend the line of "resume_state" in ./configs/options_train.json (17th line) and customize it with the specific resumed path, run

sh train.sh Deblurring/configs/options_train.json

:framed_picture: Evaluation

• Download the pre-trained models and place them in Noise-DA/checkpoints/ (Ignored if Demo & Quick Inference has been tried).

• Customize the checkpoint path (14th line) and test dataset path (23rd-24th lines) in ./configs/options_test.json, run

sh test.sh Deblurring/configs/options_test.json

• To reproduce the quantitative metrics of the RealBlur_J dataset, customize the paths of the deblurred images and ground truth images, run

python eva_deblur.py