This repository contains the unofficial PyTorch re-implementation of the paper:
Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring
Seungjun Nah, Tae Hyun Kim, Kyoung Mu Lee
In CVPR 2017

arXiv | Original Torch code

Note: This code is for personal project. Details of the implementation may different from the original code. Should you be making use of this repository for your own research, please take it with care.

Performance on GOPRO test dataset (PSNR/SSIM)

* Original paper used MATLAB implementation whereas this repo used skimage implementation.

git clone https://github.com/jinsuyoo/DeepDeblur-PyTorch.git
cd DeepDeblur-PyTorch

• Linux (tested on Ubuntu 16.04 LTS)
• Anaconda
• Python 3.7
• PyTorch 1.5.0
• TensorFlow (used for Tensorboard visualization)

and the Python dependencies listed in requirements.txt file.

To install, please run the following commands.

conda create -n DeepDeblur python=3.7
conda activate DeepDeblur
conda install pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=10.2.89 -c pytorch
pip install -r requirements.txt

Download GOPRO dataset from here.

The data folder should be organized as,

/datasets/GOPRO_Large

python train.py

python test.py

Total G loss	Content loss	Adversarial loss	D loss

PSNR on validation set (same as test set in this research)

Augmentation code borrows from the original code.