Database and VQA codes for the following paper:
@article{yixuan2023cfvqa,
title={Perceptual Quality Assessment of Face Video Compression: A Benchmark and An Effective Method},
author={Yixuan Li, Bolin Chen, Baoliang Chen, Meng Wang, and Shiqi Wang},
journal={arXiv preprint arXiv:2304.07056},
year={2023}
}
- The first large-scale subjective video quality assessment for compressed face videos. Spatial and temporal distortions, along with traditional and generative distortions for face videos are included and studied.
- The first full-reference face video quality assessment method correlating well with human perceptual quality. Free of training or fine-tuning.
In total, 3240 distorted face video sequences derived from 135 reference face videos are included in our dataset. The download link of the CFVQA dataset is
[Due to copyright issues, some videos from Avcocado.com are eliminated from the dataset and are replaced by copywrite-free data.]
Please contact YixuanLI (yixuanli423@gmail.com) for the download password, questions or comments. （If you need the face videos with larger spatial resolution and longer duration in the MP4 format, also can contact me.）
(The onedrive link might be invalid from 2025.1. If so, please contact me.) Please cite the paper if you use any part of the videos or data supplied in this project page.
The source face videos were obtained from two copyright-free website Pexel(https://www.pexels.com/videos/) and Mixkit(https://mixkit.co/), where the videos are permitted to use, modify and redistribute freely without attribution.
**Please note that the videos from mixkit.com are only free of non-commercial use, for commercial usage, please refer to the videos from the pexels.com. (If you need to know the face videos free of commercial usage, please contact me.)
Compressed videos in this database are provided as *.yuv files.
All the video files have planar YUV 4:2:0 format and contain no headers. The resolution for all videos is 512 x 512. The 8-bit Y component of the first frame, followed by the 8-bit U and 8-bit V component of the first frame. Frames are concatenated to form sequence files.
Here are some examplar compressed video clips contained in the CFVQA dataset.
1.mp4
110.mp4
The file naming convention of this database has the pattern of "num_codec_qp.yuv" or "num.yuv".
- "num" is an interger that denotes the reference video sequence that was used to create the compressed video. There are totally 135 reference videos in this database, so the patterns of "num" is from 1 to 135.
- "100.yuv" - The 100th reference face video clip in this dataset.
- "codec" means the video compression codec used for generating compressed video sequence. The patterns used for them are as follows:
- "vvc" - Versatile Video Coding (VVC)
- "rl" - reduced resolution coding
- "dvc" - DVC: An End-to-end Deep Video Compression Framework
- "rlvc" - Learning for Video Compression with Recurrent Auto-Encoder and Recurrent Probability Model
- "fomm" - First Order Motion Model for Image Animation
- "cfte" - Beyond Keypoint Coding: Temporal Evolution Inference with Compact Feature Representation for Face Video Compression
- "qp" is qp value that denotes the quantization parameter.
For instance, for the "100.yuv" sequence:
- "100.yuv" is a reference video.
- "100_dvc_22.yuv" is a compressed video obtained by compressing the reference video with the codec dvc at qp22.
- "100_rlvc_27.yuv" is a compressed video obtained with the codec rlvc with lambda=1024.
Specially, in the fomm and cfte folder: "100_fomm_22.yuv" denotes a compressed video obtained with fomm model with 10 key points, and the first frame is compressed by VVC in the QP level of 22. The same convention is for CFTE-compressed videos.
The implementation details of the employed video compression codecs can be found in the supplementary file. The adopted implementations are listed as followed.
VVC : https://vcgit.hhi.fraunhofer.de/jvet/VVCSoftware_VTM/-/tree/VTM-11.0
DVC: https://github.com/RenYang-home/OpenDVC
RLVC: https://github.com/RenYang-home/RLVC
FOMM: https://github.com/AliaksandrSiarohin/first-order-model
- Python==3.7
- Pytorch==1.8.0
- numpy
- matplotlib
- scipy
- Pillow
-
Resnet for face recognition: frmodels.py from https://github.com/Tencent/TFace/tree/quality
The pretrained parameters: /weights/IR_50_MS1M.pth
-
Multi-level spatial feature fusion parameters: /weights/DISTS.pt
Please download via https://portland-my.sharepoint.com/:u:/g/personal/yixli5-c_my_cityu_edu_hk/Ed6ZYYUpL49LurU8ac7mQjEBxB_2skaRmF5beCM1gAwAYg?e=Orbmy7 and put it in the weights folder.
- Unzip the Reference and VVC folders.
python favor_evaluation.py
28vvc.mp4
28rl.mp4
28dvc.mp4
28rlvc.mp4
28fomm.mp4
28cfte.mp4
Copyright (c) 2023 The City University of Hong Kong
All rights reserved.Permission is hereby granted, without written agreement and without license or royalty fees, to use, copy, and distribute this database (the images, the results, and the source files) and its documentation for non-commercial research and educational purposes only, provided that the copyright notice in its entirety appear in all copies of this database, and the original source of this database, all authors at The City University of Hong Kong, is acknowledged in any publication that reports research using this database.
IN NO EVENT SHALL THE CITY UNIVERSITY OF HONG KONG BE LIABLE TO ANY PARTY FOR DIRECT, INDIRECT, SPECIAL, INCIDENTAL, OR CONSEQUENTIAL LOSS OR DAMAGES ARISING OUT OF OR IN CONNECTION WITH THE USE OF OR THE INABILITY TO USE THIS DATABASE, ITS DOCUMENTATION AND INFORMATION, EVEN IF THE CITY UNIVERSITY OF HONG KONG HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH LOSS OR DAMAGE.
THE CITY UNIVERSITY OF HONG KONG SPECIFICALLY DISCLAIMS ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, ACCURACY OR COMPLETENESS AND NON INFRINGEMENT OF INTELLECTUAL PROPERTY RIGHTS OF ANY PARTY. THE DATABASE PROVIDED HEREUNDER IS ON AN "AS IS" BASIS, AND THE CITY UNIVERSITY OF HONG KONG HAS NO OBLIGATION TO PROVIDE MAINTENANCE, SUPPORT, UPDATES, ENHANCEMENTS, OR MODIFICATIONS.