4K Video Super-Resolution Dataset | BVI-DVC-SR
The dataset contains 200 4K video sequences for each one of six deep-learning-based super-resolution methods. They correspond to the original 200 videos in BVI-DVC (https://fan-aaron-zhang.github.io/BVI-DVC/). The 1080p versions of these videos have been upscaled to 4K resolution.
The dataset contains a diverse set of upscaled videos, employing a complete set of modern super-resolution methods. It allows training several purpose models, including Quality Assessment, Super-Resolution and Super-Resolution Detection.
The dataset contains around 1,000 videos, each with 64 frames (some videos have not been uploaded for licensing issues). The following methods have been used:
- RVRT https://github.com/JingyunLiang/RVRT
- BasicVSR https://ckkelvinchan.github.io/projects/BasicVSR/
- Real-BasicVSR https://github.com/ckkelvinchan/RealBasicVSR
- SwinIR Classical https://github.com/JingyunLiang/SwinIR
- SwinIR Real https://github.com/JingyunLiang/SwinIR
- Real-ESRGAN https://github.com/xinntao/Real-ESRGAN
Upcaling has been performed with the following repositories:
- https://github.com/JingyunLiang/RVRT
- https://github.com/open-mmlab/mmagic
- https://github.com/xinntao/Real-ESRGAN
- https://github.com/JingyunLiang/SwinIR
The BVI-DVC dataset contains downscaled videos from the original 4K counterparts. The 1080p versions have been used to get the final versions. For super-resolution methods that do not allow x2 upscaling, we performed x4 and then downscaled them using bicubic interpolation.