PyTorch implementation of Benchmark Dataset and Effective Inter-Frame Alignment for Real-World Video Super-Resolution

Some samples of proposed MVSR4× dataset.

• Dataset

  • we use an ultra-wide camera with a 13mm-equivalent lens and a telephoto camera with a 90mm-equivalent lens to capture LR and HR sequences, respectively.

• Download

  • Our MVSR4× dataset can be downloaded from this link with password 'uygh'.

Overall pipeline of proposed MultiAdaSTN.

• Prerequisites
  • Python $\geqslant$ 3.7, PyTorch $\geqslant$ 1.8 and CUDA $\geqslant$ 11.0.
  • OpenCV, NumPy, tqdm, lpips, scikit-image, mmcv and tensorboardX.

• For RealVSR dataset images (×2), modify dataroot in train_x2.sh and then run:

  sh train_x2.sh

• For MVSR4× dataset images (×4), modify dataroot in train_x4.sh and then run:

  sh train_x4.sh

• For RealVSR dataset images (×2), modify dataroot in test_x2.sh and then run:

  sh test_x2.sh

• For MVSR4× dataset images (×4), modify dataroot in test_x4.sh and then run:

  sh test_x4.sh

• You can specify which GPU to use by --gpu_ids, e.g., --gpu_ids 0,1, --gpu_ids 3, --gpu_ids -1 (for CPU mode). In the default setting, No.0 GPU is used.
• You can refer to options for more arguments.

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

@InProceedings{Wang_2023_CVPR,
    author    = {Wang, Ruohao and Liu, Xiaohui and Zhang, Zhilu and Wu, Xiaohe and Feng, Chun-Mei and Zhang, Lei and Zuo, Wangmeng},
    title     = {Benchmark Dataset and Effective Inter-Frame Alignment for Real-World Video Super-Resolution},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2023},
    pages     = {1168-1177}
}

This repo is built upon the framework of CycleGAN, and we borrow some code from RealVSR and BasicVSR++, thanks for their excellent work!