We developed LEHA-CVQAD dataset to evaluate full-reference and no-reference video quality metrics. Here we share the open part of the whole compression artifacts dataset (1,962 out of 6,240 videos). The hidden part is only available to benchmark-support personnel for testing metric performance. All videos are of mostly FullHD resolution, YUV420, and 10-15 seconds duration. Fps values are 24, 25, 30, 39, 50, and 60. Subjective quality scores are also provided in csv file. The higher the score is the better is the quality. To study more about the subjective quality evaluation procedure of our benchmark, you can visit the FAQ section at Subjectify.us. In total, in the full dataset there are 186 different video codecs, 5+ compression standards (including H.264/AVC, H.265/HEVC, H.266/VVC, AV1, VP9), 6,000+ compressed streams, 2M+ subjective score, 15,000+ subjective assessors, and various content, including UGC and screen content.

🏆 Leaderboard of more than 100 metrics on LEHA-CVQAD dataset: MSU Video Quality Metrics Benchmark page.

⬇️ Download

Download LEHA-CVQAD with

from huggingface_hub import snapshot_download
snapshot_download(repo_id="msm1rnov/LEHA-CVQAD", repo_type="dataset", local_dir="PATH/TO/PLACE/THE/DATASET")

You can also download all videos by wget: from the dataset folder on Synology

wget https://titan.gml-team.ru:5003/fsdownload/qrtox2FLW/Compressed_and_GT_videos.zip https://titan.gml-team.ru:5003/fsdownload/qrtox2FLW/Compressed_and_GT_videos.zip

🔎 Dataset Structure

Subjective_scores_and_videos_info.csv contains subjective scores (MOS, Bradley-Terry, ELO) for each compressed video. Each distorted video beside its subjective quality has the following characteristics:
- name of the original (pristine) video
- codec used for encoding
- codec standard (avc, hevc, vvc, av1, ...)
- target bitrate or crf
- bitrate range (high, mid, low)
- original video resolution
- original video fps
Metrics_scores.csv contains 100+ VQA metrics values on our dataset and can be used to calculate VQA metrics correlations with subjective scores
60 folders with compressed and GT videos, each of which include 1 reference videos (GT), which is required to test full-reference metrics, and many distorted videos (compressed), grouped by encoding preset:
- Each distorted video has the following pattern: {video name}/{encoding preset}/{codec name}_{crf or bitrate}.mp4
- Each reference video has the following pattern: {video name}/GT.mp4

👨🏻‍💻 Correlation Calculation for MOS

The following pipeline should be applied only to calculate correlation between metrics scores and MOS subjective scores.

Just apply single correlation coefficient to the whole list of MOS subjective scores and metrics scores.

👨🏻‍💻 Correlation Calculation for BT and ELO

The following pipeline should be applied only to calculate correlation between metrics scores and BT and ELO subjective scores.

There are 59 different original (pristine) videos, as well as several encoding presets in the dataset. Please pay attention: It is required to calculate the correlation coefficient (SRCC, KRCC, ...) on all of them SEPARATELY. Therefore, to get a single correlation for the whole dataset, you should use Fisher Z-transform to average group correlations weighted proportionally to group size as follows:

Iterate through 59 original videos and for each calculate correlation coefficients, as many times as the quantity of unique presets for the current video (i.e. for basketball-2021 with 2 presets fast and offline you should obtain 2 correlations)
Use the inverse hyperbolic tangent (artanh) on each value of the correlation coefficient
- Replace possible infinity with artanh(0.99)
Apply weighted arithmetic mean to obtained values. For example, if SROCC_1 is the spearman correlation counted for the group of samples of size Size_1, SROCC_2 is the spearman correlation counted for the group of samples of size Size_2, then the final correlation have to be counted as (SROCC_1 * SIZE_1 + SROCC_2 * SIZE_2)/(SIZE_1 + SIZE_2).
Calculate the hyperbolic tangent (tanh) of the weighted mean
- Take the absolute value of it and replace 0.99 with 1
The obtained value represents the correlation between your method scores and the subjective scores on our dataset.

Script to calculate metrics correlations with subjective scores (BT and ELO) is provided in the GitHub repo: https://github.com/msu-video-group/MSU_VQM_Compression_Benchmark

🛠️ Encoding and Decoding

To encode videos we used the following command:

    ffmpeg −f rawvideo −vcodec rawvideo −s {width}x{height} −r {FPS} −pix_fmt yuv420p −i {video name}.yuv −c:v libx265 −x265−params "lossless =1:qp=0" −vsync 0 {video name}.mp4

To decode the video back to YUV you can use:

    ffmpeg -i {video name}.mp4 -pix_fmt yuv420p -vcodec rawvideo -f rawvideo {video name}.yuv

To convert the encoded video to the set of PNG images you can use:

    ffmpeg -i {video name}.mp4 {frames dir}/frame_%05d.png

📚 Citation

If you find this dataset useful, please consider giving a like ❤️ and citation

@inproceedings{10.1145/3746027.3758303,
    author = {Gushchin, Aleksandr and Smirnov, Maksim and Vatolin, Dmitriy S. and Antsiferova, Anastasia},
    title = {LEHA-CVQAD: Dataset To Enable Generalized Video Quality Assessment of Compression Artifacts},
    year = {2025},
    publisher = {Association for Computing Machinery},
    url = {https://doi.org/10.1145/3746027.3758303},
    doi = {10.1145/3746027.3758303},
    booktitle = {Proceedings of the 33rd ACM International Conference on Multimedia},
    pages = {13405–13412},
    series = {MM '25}
}

🌐 Contact

The CMC MSU Graphics and Media Lab hosts the dataset.

If you have any question regarding the usage of LEHA-CVQAD, please feel free to contact us via [email protected]

Downloads last month: 89