arXiv Paper Version

Pytorch implementation of the paper DeblurGAN: Blind Motion Deblurring Using Conditional Adversarial Networks.

Our network takes blurry image as an input and procude the corresponding sharp estimate, as in the example:

The model we use is Conditional Wasserstein GAN with Gradient Penalty + Perceptual loss based on VGG-19 activations. Such architecture also gives good results on other image-to-image translation problems (super resolution, colorization, inpainting, dehazing etc.)

• NVIDIA GPU + CUDA CuDNN (CPU untested, feedback appreciated)
• Pytorch

Download weights from Dropbox . Note that during the inference you need to keep only Generator weights.

Put the weights into

/.checkpoints/experiment_name

To test a model put your blurry images into a folder and run:

python test.py --dataroot /.path_to_your_data --model test --dataset_mode single --learn_residual

Download dataset for Object Detection benchmark from Google Drive

The code used to create the paper contains a bug in PSNR calculation, found by Huaijin Chen from Rice university, results recalculation are in progress.

If you find our code helpful in your research or work please cite our paper.

@article{DeblurGAN,
  title = {DeblurGAN: Blind Motion Deblurring Using Conditional Adversarial Networks},
  author = {Kupyn, Orest and Budzan, Volodymyr and Mykhailych, Mykola and Mishkin, Dmytro and Matas, Jiri},
  journal = {ArXiv e-prints},
  eprint = {1711.07064},
  year = 2017
}

Code borrows heavily from pix2pix. The images were taken from GoPRO test dataset - DeepDeblur