DeepDeblur

Make sure to clone with submodules

This is a repository for the 10701 project "Blur Kernel Estimation and Tracking via GANs". The

Make sure to clone with submodules

git clone --recurse-submodules https://github.com/Harsharma2308/DeepDeblur.git

Anaconda environment

conda create --name py37 python=3.7

Prerequisites

• NVIDIA GPU + CUDA CuDNN
• Pytorch

Dependencies

pip install -r requirements.txt

Training Generators

cd DeepDeblur/Blind-Image-Deconvolution-using-Deep-Generative-Priors/scripts
python svhn_train.py
python blur_train.py
python DCGAN_train.py

Inference

cd gan_train
Run `deblurring_SVHN_wo_init.py`  for running the algorithm without the modified initialisation of latent space.
Run `deblurring_SVHN_with_init.py`  for running the algorithm with initialisation of latent space.

Install dependencies for training generators

pip install -r requirements.txt

Datasets

Download SVHN train dataset

cd gan_train/data
wget http://ufldl.stanford.edu/housenumbers/train_32x32.mat

Download Blurryvideo dataset

blurVideoSVHN

Create Blurkernel dataset

cd Blind-Image-Deconvolution-using-Deep-Generative-Priors/blur_data_generation
matlab -nodisplay -nodesktop -r "run blur_data_generate.m"

Hyper parameters used

REGULARIZORS = [0.01 , 0.01] alpha = 1.0 (for Algorithm 2) NOISE_STD = 0.01 STEPS = 6000