🏆 [CVPRW 2024] COVER: A Comprehensive Video Quality Evaluator.

🏆 🥇 Winner solution for Video Quality Assessment Challenge at the 1st AIS 2024 workshop @ CVPR 2024

Official Code for [CVPR Workshop 2024] Paper "COVER: A Comprehensive Video Quality Evaluator". Official Code, Demo, Weights for the Comprehensive Video Quality Evaluator (COVER).

• 17 June 2024: We presented our solution in the AIS 2024 workshop in CVPR 2024! See the photo here.
• 29 May 2024: We create a space for COVER on Hugging Face, with ZeroGPU enabled!
• 09 May 2024: We upload the Code of COVER.
• 12 Apr 2024: COVER has been accepted by CVPR Workshop2024.

Introduction

• Existing UGC VQA models strive to quantify quality degradation mainly from the technical aspect, with a few considering aesthetic or semantic aspects, but no model has addressed all three aspects simultaneously.
• The demand for high-resolution and high-frame-rate videos on social media platforms presents new challenges for VQA tasks, as they must ensure effectiveness while also meeting real-time requirements.

COVER: A Completely Blind Video Quality Evaluator

This inspires us to develop a comprehensive and efficient model for the UGC VQA task

COVER contains three branches: (1) Semantic, (2) Aesthetic, and (3) Technical braches for comprehensive assessment of video quality. We use the semantic branch to gate the other two branches to conduct feature interactions. The final quality score is aggregated on top of the three scores.

Results

Results highlight on the AIS 2024 UGC Video Quality Challenge:

• The run time is measured on an NVIDIA A100 GPU. A clip of 30 frames of 4K resolution 3840×2160 is used as input.
• Inference time break-down is:

Results on other UGC datasets

Ablations

Install

The repository can be installed via the following commands:

git clone https://github.com/vztu/COVER 
cd COVER 
pip install -e . 
mkdir pretrained_weights 
cd pretrained_weights 
wget https://github.com/vztu/COVER/blob/release/Model/COVER.pth
cd ..

Evaluation: Judge the Quality of Any Image and Video

Try on Demos

You can run a single command to judge the quality of the demo videos in comparison with videos in VQA datasets.

    python evaluate_one_image.py -i ./demo/image_1.mp4

or

    python evaluate_one_video.py -v ./demo/video_2.mp4

Or choose any video you like to predict its quality:

    python evaluate_one_video.py -v $YOUR_SPECIFIED_VIDEO_PATH$

Outputs

The script can directly score the video's overall quality (considering all perspectives).

    python evaluate_one_video.py -v $YOUR_SPECIFIED_VIDEO_PATH$

The final output score is the sum of all perspectives.

Evaluate on a Exsiting Video Dataset

    python evaluate_one_dataset.py -in $YOUR_SPECIFIED_DIR$ -out $OUTPUT_CSV_PATH$

Evaluate on a Set of Unlabelled Videos

    python evaluate_a_set_of_videos.py -in $YOUR_SPECIFIED_DIR$ -out $OUTPUT_CSV_PATH$

The results are stored as .csv files in cover_predictions in your OUTPUT_CSV_PATH.

Please feel free to use COVER to pseudo-label your non-quality video datasets.

Data Preparation

We have already converted the labels for most popular datasets you will need for Blind Video Quality Assessment, and the download links for the videos are as follows:

:book: LSVQ: Github

:book: KoNViD-1k: Official Site

:book: LIVE-VQC: Official Site

:book: YouTube-UGC: Official Site

(Please contact the original authors if the download links were unavailable.)

After downloading, kindly put them under the ../datasets or anywhere but remember to change the data_prefix respectively in the config file.

Training: Adapt COVER to your video quality dataset!

Now you can employ head-only/end-to-end transfer of COVER to get dataset-specific VQA prediction heads.

    python transfer_learning.py -t $YOUR_SPECIFIED_DATASET_NAME$

For existing public datasets, type the following commands for respective ones:

• python transfer_learning.py -t val-kv1k for KoNViD-1k.
• python transfer_learning.py -t val-ytugc for YouTube-UGC.
• python transfer_learning.py -t val-cvd2014 for CVD2014.
• python transfer_learning.py -t val-livevqc for LIVE-VQC.

As the backbone will not be updated here, the checkpoint saving process will only save the regression heads. To use it, simply replace the head weights with the official weights COVER.pth.

We also support end-to-end fine-tune right now (by modifying the num_epochs: 0 to num_epochs: 15 in ./cover.yml). It will require more memory cost and more storage cost for the weights (with full parameters) saved, but will result in optimal accuracy.

Visualization

WandB Training and Evaluation Curves

You can be monitoring your results on WandB!

Acknowledgement

Thanks for every participant of the subjective studies!

Citation

Should you find our work interesting and would like to cite it, please feel free to add these in your references!

%AIS 2024 VQA challenge
@article{conde2024ais,
  title={AIS 2024 challenge on video quality assessment of user-generated content: Methods and results},
  author={Conde, Marcos V and Zadtootaghaj, Saman and Barman, Nabajeet and Timofte, Radu and He, Chenlong and Zheng, Qi and Zhu, Ruoxi and Tu, Zhengzhong and Wang, Haiqiang and Chen, Xiangguang and others},
  journal={arXiv preprint arXiv:2404.16205},
  year={2024}
}

%cover
@InProceedings{cover2024cpvrws,
    author    = {He, Chenlong and Zheng, Qi and Zhu, Ruoxi and Zeng, Xiaoyang and Fan, Yibo and Tu, Zhengzhong},
    title     = {COVER: A Comprehensive Video Quality Evaluator},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2024},
    pages     = {5799-5809}
}