Deep Video Deblurring Using Sharpness Features from Exemplars (Submitted to IEEE TIP, 2020)

by Xinguang Xiang, Hao Wei, Jinshan Pan

Dependencies and Installation

• Python 3 (Recommend to use Anaconda)
• PyTorch0.4.1
• Linux (Tested on Ubuntu 18.04)
• numpy
• tqdm
• imageio
• matplotlib

Dataset Preparation

We use the GOPRO_Su dataset to train our models. You can download it from here and put the dataset into 'train_dataset/'. The dataset should be organized in the following form:

|--dataset name  |--train   |--video 01    |--input     |--frame 01     |--frame 02     |...    |--GT     |--frame 01     |--frame 02     |...   |--video 02   ...  |--val  |--test

Training

• Download the FlowNet pretrained model from Baidu Drive (password:2gca) and put it into 'pretrained_model/'.
  
• Prepare the dataset same as above form.
  
• Start to train the model. Hyper parameters such as batch size, learning rate, epoch number can be tuned through command line:

python main.py --batch_size 4 --patch_size 256 --lr 1e-4 --epochs 500 --save_models

Testing

• Download our pretrained model from Baidu Drive (password:2gca) .
• The testing command is shown as follows:

python main.py --pre_train model_path --test_only

Testing your own data

• Put your testing data into 'inference/' which should be organized the same as our given examplars.
• The testing command is shown as follows:

python inference.py

Citation

@article{xiang2020DVD-SFE,
title={Deep Video Deblurring Using Sharpness Features from Exemplars},
author={Xiang, xinguang and Wei, Hao and Pan, jinshan},
journal={IEEE Transactions on Image Processing},
year={2020}
}