FVQ-20K Dataset

Overview

• FVQ-20K is an in-the-wild face video quality assessment (FVQA) dataset, which contains 20,000 face videos with MOS annotations.
  
• The FVQ-20K dataset is divided into training, validation, and test sets with a ratio of 80% : 5% : 15%.

Data Structure

FVQ-20K
│
├── train
│   ├── labels.txt
│   └── videos
│        ├── *.mp4
│        └── ...
├── val
│   ├── labels.txt
│   └── videos
│        ├── *.mp4
│        └── ...
└── test
    ├── labels.txt
    └── videos
         ├── *.mp4
         └── ...

• labels.txt contains video names and their corresponding MOS scores (ranging from 0 to 100).

Citation

If you use this dataset, please consider citing

@inproceedings{wu2025fvq,
  title={FVQ: A Large-Scale Dataset and an LMM-based Method for Face Video Quality Assessment},
  author={Wu, Sijing and Li, Yunhao and Xu, Ziwen and Gao, Yixuan and Duan, Huiyu and Sun, Wei and Zhai, Guangtao},
  booktitle={Proceedings of the 33rd ACM International Conference on Multimedia},
  pages={6928--6937},
  year={2025}
}

Contact

• Sijing Wu (wusijing@sjtu.edu.cn)