Home

Awesome

Dataset and Code for RealVSR

Real-world Video Super-resolution: A Benchmark Dataset and A Decomposition based Learning Scheme
Xi Yang, Wangmeng Xiang, Hui Zeng and Lei Zhang
International Conference on Computer Vision, 2021.

Dataset

The dataset is hosted on Google Drive and Baidu Drive (code: 43ph). Some example scenes are shown below.

dataset_samples

The structure of the dataset is illustrated below.

FileDescription
GT.zipAll ground truth sequences in RGB format
LQ.zipAll low quality sequences in RGB format
GT_YCbCr.zipAll ground truth sequences in YCbCr format
LQ_YCbCr.zipAll low quality sequences in YCbCr format
GT_test.zipGround truth test sequences in RGB format
LQ_test.zipLow Quality test sequences in RGB format
GT_YCbCr_test.zipGround truth test sequences in YCbCr format
LQ_YCbCr_test.zipLow Quality test sequences in YCbCr format
videos.zipOriginal videos (> 500 LR-HR pairs here)

Code

Dependencies

Installation

# Create a new anaconda python environment (realvsr)
conda create -n realvsr python=3.7 -y

# Activate the created environment
conda activate realvsr

# Install dependencies
pip install -r requirements.txt

# Bulid the DCN module
cd codes/models/archs/dcn
python setup.py develop

Training

Modify the configuration files accordingly in codes/options/train folder and run the following command (current we did not implement distributed training):

python train.py -opt xxxxx.yml

Testing

Test on RealVSR testing set sequences:

Modify the configuration in test_RealVSR_wi_GT.py and run the following command:

python test_RealVSR_wi_GT.py

Test on real-world captured sequences:

Modify the configuration in test_RealVSR_wo_GT.py and run the following command:

python test_RealVSR_wo_GT.py

Pre-trained Models

Some pretrained models could be found on Google Drive and Baidu Drive (code: n1n0).

License

This project is released under the Apache 2.0 license.

Citation

If you find this code useful in your research, please consider citing:

@article{yang2021real,
  title={Real-world Video Super-resolution: A Benchmark Dataset and A Decomposition based Learning Scheme},
  author={YANG, Xi and Xiang, Wangmeng and Zeng, Hui and Zhang, Lei},
  journal=ICCV,
  year={2021}
}

Acknowledgement

This implementation largely depends on EDVR. Thanks for the excellent codebase! You may also consider migrating it to BasicSR.