nickossaf/Screen-Content-Video

Screen Content Video Dataset

Subjective dataset of screen-content videos for video quality assessment. It contains 100 reference sequences and 1,600 distorted sequences, with one subjective score for every distorted sequence.

The dataset is described in:

• Project page: videoprocessing.github.io/screen-content-dataset
• Paper: Nikolai Safonov, Mikhail Rakhmanov, and Dmitriy Vatolin. “Screen Content Video Dataset and Benchmark.” Proceedings of the 33rd ACM International Conference on Multimedia (MM ’25), 2025. DOI: 10.1145/3746027.3758306. Pages 13428–13434.

Screen sharing is a common feature of video conferencing, and limited bandwidth, packet loss, and compression change how that content looks. Most quality metrics were designed for natural scenes. This dataset is a subjective benchmark for screen content: slides, code editors, terminals, web pages, spreadsheets, charts, mixed interfaces, and games, including both static and dynamic sequences.

There is no train, validation, or test split in this release. Every sequence is in one pool.

What is in this folder

Item	Count
Reference videos (orig.mp4)	100
Distorted videos	1,600
MP4 files	1,700
Subjective scores	1,600
Total size of the videos and the CSV	4,440,135,620 bytes (4.14 GiB)

subjective_100_scores.csv
1/
  orig.mp4
  av1_1.mp4
  h264_1.mp4
  din_1.mp4
  din_3.mp4
  dis_1.mp4
  dis_3.mp4
  meet_2.mp4
  teams_2.mp4
  tel_1.mp4
  tel_3.mp4
  tube_1.mp4
  tube_3.mp4
  voov_1.mp4
  voov_3.mp4
  zoom_1.mp4
  zoom_3.mp4
2/
  ...
100/
  ...

Folders 1 through 100 are the reference sequences. Each folder has the same 17 files. The filename without .mp4 is the method id used in the score table. orig is the reference and has no score row.

Scores

subjective_100_scores.csv has 1,600 rows and three columns:

Column	Type	Meaning
video	integer 1–100	Reference sequence. Matches the folder name.
method	string	Distortion id. Matches the MP4 filename without the extension.
score	float	Subjective quality score for that distorted video.

Every sequence has all 16 methods, and every method has all 100 sequences. In this file the scores run from 0.0 to 8.385516169545436.

The project page says the scores come from crowdsourced pairwise comparisons. Participants saw two videos side by side and chose “left”, “right”, or “can’t choose”. Each participant rated 20 pairs, including 2 hidden validation pairs made by compressing the originals at a high CRF.

This CSV does not include a text description of each distortion. The project page mentions a description column in Subjective_scores.csv; that column is not in subjective_100_scores.csv.

Method ids

The project page describes two ways the distorted videos were made: lossless screen recordings of video-conferencing applications under network conditions emulated with tc-netem, and compression with standard codecs. It names Google Meet and Microsoft Teams as the two services whose recordings are not pixel-aligned with the reference. Those recordings should be left out of full-reference metric checks. In this folder they are meet_2 and teams_2.

The other ids are stored as filenames only. This folder has no legend that maps the suffix _1 or _3 to a bandwidth, a packet-loss rate, or a CRF. Those parameters are not invented here.

Method id	Files	Resolution in this folder	Note from the project page
av1_1	100	1535×1030	Codec-compression condition. The stored bitstream is H.264, not AV1.
h264_1	100	1535×1030	Codec-compression condition.
din_1, din_3	100 each	1535×1030	Id only.
dis_1, dis_3	100 each	1535×1030	Id only.
meet_2	100	1319×883	Google Meet. Not pixel-aligned; exclude from full-reference evaluation.
teams_2	100	1326×884	Microsoft Teams. Not pixel-aligned; exclude from full-reference evaluation.
tel_1, tel_3	100 each	1535×1030	Id only.
tube_1, tube_3	100 each	1535×1030	Id only.
voov_1, voov_3	100 each	1535×1030	Id only.
zoom_1, zoom_3	100 each	1535×1030	Id only.

Video properties measured in this folder

All 1,700 files were opened with ffprobe.

• No file has an audio stream.
• All 100 av1_1.mp4 files, and every file in sequence 1, are H.264 High 4:4:4 Predictive, pixel format yuv444p, in an MP4 container.
• orig.mp4, av1_1, h264_1, din_*, dis_*, tel_*, tube_*, voov_*, and zoom_* are 1535×1030.
• Every meet_2.mp4 is 1319×883.
• Every teams_2.mp4 is 1326×884.
• Every orig.mp4 lasts 10.000 seconds.
• References 1, 50, and 100 have 300 frames and an average frame rate of 960000/31999 (about 30 fps). Distorted files can have a different frame count. Sequence 1, teams_2.mp4, has 296 frames.

The project page describes the capture, before the timer was removed, as 1920×1080, 30 fps, 7–10 seconds, recorded with OBS Studio, lossless compression, and 4:4:4 chroma. A timer region was used to align frames and was cropped out before the subjective test. Conferencing apps can drop frames, so a distorted file may not contain every reference frame.

frames_mappings.json is described on the project page and is not in this folder. The page also names the reference orig.mkv; the reference here is orig.mp4.

How to load a few examples

The repository is a set of files, not a packaged datasets script.

from huggingface_hub import hf_hub_download
import pandas as pd

repo_id = "nickossaf/Screen-Content-Video"

scores = pd.read_csv(
    hf_hub_download(
        repo_id=repo_id,
        filename="subjective_100_scores.csv",
        repo_type="dataset",
    )
)

# one reference and two distorted versions of sequence 1
paths = [
    hf_hub_download(repo_id, name, repo_type="dataset")
    for name in ("1/orig.mp4", "1/h264_1.mp4", "1/meet_2.mp4")
]

To read the score for one of those videos:

row = scores[(scores.video == 1) & (scores.method == "h264_1")].iloc[0]
print(row.score)

License

Creative Commons Attribution 2.0 Generic (CC BY 2.0).

Citation

@inproceedings{scvqa2025,
  author    = {Nikolai Safonov and Mikhail Rakhmanov and Dmitriy Vatolin},
  title     = {Screen Content Video Dataset and Benchmark},
  booktitle = {Proceedings of the 33rd ACM International Conference on Multimedia (MM '25)},
  year      = {2025},
  pages     = {13428--13434},
  publisher = {Association for Computing Machinery},
  doi       = {10.1145/3746027.3758306},
  isbn      = {979-8-4007-2035-2/2025/10}
}

Contact

From the project page: nikolay.safonov@graphics.cs.msu.ru, mihail.rakhmanov@graphics.cs.msu.ru, dmitriy@graphics.cs.msu.ru.