README.md

Training

1. To download GoPro training and testing data, run

python download_data.py --data train-test

2. Generate image patches from full-resolution training images of GoPro dataset

python generate_patches_gopro.py 

3. To train Restormer, run

cd Restormer
./train.sh Motion_Deblurring/Options/Deblurring_Restormer.yml

Note: The above training script uses 8 GPUs by default. To use any other number of GPUs, modify Restormer/train.sh and Motion_Deblurring/Options/Deblurring_Restormer.yml

Evaluation

Download the pre-trained model and place it in ./pretrained_models/

Testing on GoPro dataset

• Download GoPro testset, run

python download_data.py --data test --dataset GoPro

• Testing

python test.py --dataset GoPro

Testing on HIDE dataset

• Download HIDE testset, run

python download_data.py --data test --dataset HIDE

• Testing

python test.py --dataset HIDE

Testing on RealBlur-J dataset

• Download RealBlur-J testset, run

python download_data.py --data test --dataset RealBlur_J

• Testing

python test.py --dataset RealBlur_J

Testing on RealBlur-R dataset

• Download RealBlur-R testset, run

python download_data.py --data test --dataset RealBlur_R

• Testing

python test.py --dataset RealBlur_R

To reproduce PSNR/SSIM scores of the paper (Table 2) on GoPro and HIDE datasets, run this MATLAB script

evaluate_gopro_hide.m 

To reproduce PSNR/SSIM scores of the paper (Table 2) on RealBlur dataset, run

evaluate_realblur.py