maxtbit / nrknet Goto Github PK

This repo contains training and evaluation code for the following paper:

Neumann Network with Recursive Kernels for Single Image Defocus Deblurring.

IEEE Computer Vision and Pattern Recognition (CVPR) 2023

conda create -n NRKNet python=3.7
conda activate NRKNet
cd ./NRKNet
pip install -r requirements.txt

Download and unzip datasets under [DATASET_ROOT]:

• DPDD dataset: Google Drive | Dropbox
• CUHK test set: Google Drive | Dropbox
• RealDOF test set: Google Drive | Dropbox
• LFDOF dataset: Download path
• RTF test set: Please contact the author (Laurent, [email protected]), he will kindly share the dataset and necessary information with you.

[DATASET_ROOT]
 ├── DPDD
 ├── RealDOF
 ├── CUHK
 ├── LFDOF
 └── RTF
 

[DATASET_ROOT] can be modified with [config.data_offset] in ./config.py.

Train the NRKNet with different training datasets (DPDD | LFDOF).

# Trained with DPDD
CUDA_VISIBLE_DEVICES=0 python train_DPDD.py

# Trained with LFDOF
CUDA_VISIBLE_DEVICES=0 python train_LFDOF.py

Download the pre-trained models and unzip datasets under [NRKNet-main]:

• Select the training and testing datasets in config.py.

  • 'train['train_dataset_name']': The name of a dataset to train. DPDD | LFDOF. Default: DPDD
  • 'test['dataset']': The name of a dataset to evaluate. DPDD | LFDOF | RTF| RealDOF. Default: DPDD

• Run test.py.

CUDA_VISIBLE_DEVICES=0 python test.py

• Modify the path of a re-trained model in config.py.

# From
train['resume'] = './save/NRKNet_' + train['train_dataset_name'] + '/0'

#To
train['resume'] = './save/NRKNet_' + train['train_dataset_name'] + '/1'

• Select the training and testing datasets in config.py.

• Run test.py

Open an issue for any inquiries. You may also have contact with [email protected]

@inproceedings{quan2023neumann,
  title={Neumann Network With Recursive Kernels for Single Image Defocus Deblurring},
  author={Yuhui Quan, Zicong Wu and Hui Ji},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={5754--5763},
  year={2023}
}