SOF-VSR (Super-resolving Optical Flow for Video Super-Resolution)

Figure 2. Comparison with the state-of-the-arts.

Requirements

• Python 3
• pytorch (0.4), torchvision (0.2)
• numpy, PIL
• Matlab (For PSNR/SSIM evaluation)

Datasets

We collect 145 1080P video clips from the CDVL Database for training.

We use the Vid4 dataset and a subset of the DAVIS dataset (namely, DAVIS-10) for benchmark test.

• Vid4(BaiduPan, GoogleDrive)
• DAVIS-10
  We use 10 scenes in the DAVIS-2017 test set including boxing, demolition, dive-in, dog-control, dolphins, kart-turn, ocean-birds, pole-vault, speed-skating and wings-trun.

Train & Test

[ACCV]

[TIP]

Results

Figure 3. Comparative results achieved on the Vid4 dataset. Zoom-in regions from left to right: IDNnet, VSRnet, TDVSR, our SOF-VSR, DRVSR and our SOF-VSR-BD.

Figure 4. Comparative results achieved on the DAVIS-10 dataset. Zoom-in regions from left to right: IDNnet, VSRnet, our SOF-VSR, DRVSR and our SOF-VSR-BD.

Figure 5. Visual comparison of 4x SR results. From left to right: VSRnet, TDVSR, our SOF-VSR and the groundtruth.

Citation

@InProceedings{Wang2018accv,
  author    = {Longguang Wang and Yulan Guo and Zaiping Lin and Xinpu Deng and Wei An},
  title     = {Learning for Video Super-Resolution through {HR} Optical Flow Estimation},
  booktitle = {ACCV},
  year      = {2018},
}
@Article{Wang2020tip,
  author    = {Longguang Wang and Yulan Guo and Li Liu and Zaiping Lin and Xinpu Deng and Wei An},
  title     = {Deep Video Super-Resolution using {HR} Optical Flow Estimation},
  journal   = {{IEEE} Transactions on Image Processing},
  year      = {2020},
}

Contact

For questions, please send an email to wanglongguang15@nudt.edu.cn