README.md

📖 Learning Efficient Deep Discriminative Spatial and Temporal Networks for Video Deblurring

Jinshan Pan, Long Sun, Boming Xu, Jiangxin Dong, and Jinhui Tang
IMAG Lab, Nanjing University of Science and Technology

---

This repo is a official implementation of "Learning Efficient Deep Discriminative Spatial and Temporal Networks for Video Deblurring".

DSTNet+ is an extension of DSTNet.

📜 News

• 2025.03.25: All pretrained models and visual results are available.
• 2025.03.25: The paper can be found here.
• 2025.03.14: This paper is accepted by TPAMI.
• 2024.01.08: This repo is created.

🚀 Quick Started

1. Environment Set Up

• Python 3.9, PyTorch == 1.13
• BasicSR 1.4.2
• Platforms: Ubuntu 18.04, cuda-11

git clone https://github.com/sunny2109/DSTNet-plus.git
cd DSTNet-plus
conda create -n dstnetplus python=3.9
conda activate dstnetplus
# Install dependent packages
pip install -r requirements.txt
# Install cupy
# Please make sure that the installed Cupy version matches your existing CUDA installation!
pip install cupy-cuda11x
# Install BasicSR
python setup.py develop

2. Download datasets

Used training and testing sets can be downloaded as follows:

Training Set	Pretrained model	Visual Result
GoPro	Hugging Face | Github | Baidu Cloud	Hugging Face or Baidu Cloud
DVD	Hugging Face | Github | Baidu Cloud	Hugging Face or Baidu Cloud
BSD	Hugging Face | Github | Baidu Cloud	Hugging Face or Baidu Cloud
DAVIS-2017	Hugging Face | Github | Baidu Cloud	Hugging Face or Baidu Cloud

3. Run the training code

# train DSTNetPlus on GoPro dataset
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python -m torch.distributed.launch --nproc_per_node=8 --master_port=4321 python basicsr/train.py -opt options/train/train_base_GoPro.yml --launcher pytorch

# train DSTNetPlus on DVD dataset
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python -m torch.distributed.launch --nproc_per_node=8 --master_port=4321 python basicsr/train.py -opt options/train/train_base_DVD.yml --launcher pytorch

# train DSTNetPlus on BSD dataset
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
python -m torch.distributed.launch --nproc_per_node=8 --master_port=4321 python basicsr/train.py -opt options/train/train_base_BSD1ms.yml --launcher pytorch

4. Quick inference

• Download the pretrained models.

Please download the pretrained and put it in ./checkpoints/.

• Download the testing dataset.

Please download the test dataset and put it in ./datasets/.

• Run the following commands:

python basicsr/test.py -opt options/test/test_base_GoPro.yml
cd results
python merge_full.py

• The test results will be in './results'.

👀 Results

We achieve SOTA performance on a set of blurring datasets. Detailed results can be found in the paper. All visual results of DSTNetPlus can be downloaded here.

Click to expand

• Model efficiency (PSNR vs. Runtime vs. Params)

• Quantitative evaluations
            Evaluation on GoPro dataset         Evaluation on DVD dataset
  

• Quantitative evaluations on the BSD dataset

• Quantitative evaluations on the Set8 dataset

• Deblurred results on GoPro dataset

• Deblurred results on DVD dataset

• Deblurred results on Real-world blurry frames

📧 Contact

If you have any questions, please feel free to reach us out at cs.longsun@gmail.com

📎 Citation

If you find our work helpful for your research, please consider giving a star ⭐ and citation 📝

@article{DSTNetPlus,
  title={Learning Efficient Deep Discriminative Spatial and Temporal Networks for Video Deblurring},
  author={Pan, Jinshan and Sun, Long and Xu Boming and Dong, Jiangxin and Tang, Jinhui},
  journal={TPAMI},
  year={2025}
}