Request access to SoccerNet Raw HQ
Complete the information below to request access. The SoccerNet team will review your request and the information you provide. If your request is accepted, a personalized Non-Disclosure Agreement will be sent to the email address associated with your Hugging Face account.
By submitting this request, you confirm that the information supplied is accurate and that you have read and accept the SoccerNet Non-Disclosure Agreement available at https://drive.google.com/file/d/1Efz8yP-baa8CtcMB7SYIID7Bn5sH7sGTB4-u1YdVRf4/view.
Log in or Sign Up to review the conditions and access this dataset content.
SoccerNet Raw HQ
SoccerNet Raw HQ provides controlled access to the original SoccerNet broadcast videos and several research-ready derivatives. The repository is intended for academic research in soccer video understanding, including action spotting, temporal localization, representation learning, video retrieval, and multimodal analysis.
The original SoccerNet release contains 500 complete broadcast games, covering six major European competitions, three seasons from 2014 to 2017, and approximately 764 hours of video. Different representations of the media are stored in separate Hugging Face branches so that users can download only the format required by their research.
Important: this repository is gated and protected by a Non-Disclosure Agreement (NDA). Access is limited to research and non-commercial use. Receiving access does not grant ownership, copyright, or any other intellectual-property right in the videos or their contents.
Access procedure
- Sign in to your personal Hugging Face account.
- Complete the gated-access form on this dataset page.
- Provide your full name, affiliation, and intended research use.
- Read and accept every NDA and copyright acknowledgement.
- Submit the request.
After submission, the request remains pending while the SoccerNet team reviews the information provided, the intended research use, and the NDA and copyright acknowledgements. A personalized PDF copy of the NDA will be prepared using the submitted information and sent to the email address associated with the requesting Hugging Face account. If the request satisfies the access conditions, access will then be granted to the requesting Hugging Face username.
Review times may vary. Before submitting, ensure that the email address associated with your Hugging Face account is current and can receive the NDA.
Access is personal and account-specific. Do not share your Hugging Face token, downloaded files, extracted frames, clips, or any other copy or derivative that would expose the protected media to another person.
Available branches and data representations
The main branch contains the dataset card and repository metadata. The media are distributed through the following branches:
The inventory below was obtained from the complete recursive Hugging Face file trees. Sizes are decimal totals reported by the Hub and include the small .gitattributes file in each branch. They are approximate planning figures and may change if a branch is updated.
| Branch | Files stored for each match | Data files | Approximate branch size | Recommended use |
|---|---|---|---|---|
videos-HQ |
1_HQ.mkv, 2_HQ.mkv, and video.ini |
1,500 | 2.19 TB | Work requiring the highest available quality or the original recordings. |
videos-720p |
1_720p.mkv and 2_720p.mkv |
1,000 | 995 GB | General video understanding, action spotting, tracking, calibration, and efficient training. |
videos-224p |
1_224p.mkv and 2_224p.mkv |
1,000 | 182 GB | Lightweight baselines, rapid experimentation, and lower-storage workflows. |
frames-720p-2fps |
1.tar and 2.tar |
1,000 | 765 GB | Frame-based models and feature extraction without repeatedly decoding full videos. |
clips-720p-10s |
1.tar and 2.tar |
1,000 | 1.68 TB | Clip-level classification, representation learning, and action-centered experiments. |
clips-720p-10s-zip |
1.zip and 2.zip |
1,000 | 1.68 TB | Frame-based clip models and training pipelines that expect extracted images. |
The data-file count excludes .gitattributes. Every representation covers 500 matches, normally with one object for the first half and one for the second half. The two archive files in a game directory therefore represent the two match halves, not two complete matches. The videos-HQ count comprises exactly 1,000 MKV files and 500 video.ini files.
What each representation contains
videos-HQ
This branch contains the original high-quality broadcast representation, split into first- and second-half Matroska files. Each match normally has:
1_HQ.mkv
2_HQ.mkv
video.ini
The video.ini file records the temporal information used to locate the match halves in the original source. This is the largest full-video branch and should be selected only when the original quality is necessary.
videos-720p
This branch contains temporally trimmed, resized videos. Every match has two 720p MKV files, one per half:
1_720p.mkv
2_720p.mkv
The standard SoccerNet 720p representation is 1280 × 720 at 25 frames per second.
videos-224p
This branch contains the same two-half organization at a lower spatial resolution:
1_224p.mkv
2_224p.mkv
The standard SoccerNet 224p representation is 398 × 224 at 25 frames per second.
frames-720p-2fps
This branch contains 720p frames pre-extracted at 2 frames per second. Each match has one tar archive per half:
1.tar
2.tar
Use this branch when a pipeline consumes images rather than decoding videos at training time. Preserve the enclosing competition/season/game path when extracting the archives.
clips-720p-10s
This branch contains pre-extracted 10-second, 720p clips centered around annotated actions of interest. The clips for each match half are grouped into tar archives:
1.tar
2.tar
The outer repository objects are tar files; list an archive after download to inspect its clip member names and internal paths.
clips-720p-10s-zip
This branch contains the frame representation of the same 10-second, action-centered clips. The frames for each match half are grouped into ZIP archives:
1.zip
2.zip
The current Hub tree uses .zip, not .tar, for this branch. List an archive after download to inspect its frame member names and internal paths.
Standard SoccerNet organization
Where applicable, matches follow the standard SoccerNet hierarchy:
<league>/<season>/<game>/
with game directory names following the convention:
<date> - <time> <home team> <score> <away team>
The recursive branch trees contain the same six top-level competition directories:
england_epl
europe_uefa-champions-league
france_ligue-1
germany_bundesliga
italy_serie-a
spain_laliga
Each competition is divided into 2014-2015, 2015-2016, and 2016-2017. Across those paths, each branch contains exactly 500 game directories.
Download instructions
1. Install the Hugging Face client
python -m pip install --upgrade huggingface_hub
2. Authenticate
After your request has been approved, create a Hugging Face token with read access and authenticate your machine:
hf auth login
Do not publish, commit, or share your token.
3. Inspect the download first
The branches can be large. Use a dry run before downloading an entire representation:
hf download SoccerNet/SoccerNet_raw_HQ \
--repo-type dataset \
--revision videos-720p \
--dry-run
4. Download one complete branch
Choose the branch that matches your workflow:
hf download SoccerNet/SoccerNet_raw_HQ \
--repo-type dataset \
--revision videos-720p \
--local-dir ./SoccerNet_raw_HQ/videos-720p
Replace videos-720p in both places with any supported branch:
videos-HQ
videos-720p
videos-224p
frames-720p-2fps
clips-720p-10s
clips-720p-10s-zip
Download each representation into a different local directory. This avoids mixing files from independent branches.
Downloading a complete branch may require between approximately 182 GB and 2.19 TB of free space, plus temporary or extracted-copy space. For the archive branches, plan for substantially more space if the archives will be retained after extraction. Prefer a selective download unless the entire representation is required.
5. Download selected files or paths
After inspecting the branch in the Hugging Face file browser, pass one or more repository-relative paths to hf download:
hf download SoccerNet/SoccerNet_raw_HQ \
"<repository-relative-path>" \
--repo-type dataset \
--revision videos-720p \
--local-dir ./SoccerNet_raw_HQ/videos-720p
You can also filter a download with --include and --exclude glob patterns. Run hf download --help for the options supported by your installed client.
For example, download one competition and season from the 720p branch:
hf download SoccerNet/SoccerNet_raw_HQ \
--repo-type dataset \
--revision videos-720p \
--include "england_epl/2015-2016/**" \
--local-dir ./SoccerNet_raw_HQ/videos-720p
Download only the second-half 224p videos:
hf download SoccerNet/SoccerNet_raw_HQ \
--repo-type dataset \
--revision videos-224p \
--include "**/2_224p.mkv" \
--local-dir ./SoccerNet_raw_HQ/videos-224p
The repository is a collection of large media files rather than a dataset-builder script. Use hf download or snapshot_download for retrieval. A Git clone is generally less convenient for these multi-hundred-gigabyte and multi-terabyte branches.
Python
Use snapshot_download to retrieve a complete branch programmatically:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="SoccerNet/SoccerNet_raw_HQ",
repo_type="dataset",
revision="videos-720p",
local_dir="./SoccerNet_raw_HQ/videos-720p",
)
To select only part of a branch, add allow_patterns and/or ignore_patterns:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="SoccerNet/SoccerNet_raw_HQ",
repo_type="dataset",
revision="frames-720p-2fps",
local_dir="./SoccerNet_raw_HQ/frames-720p-2fps",
allow_patterns=["<path-or-glob-pattern>"],
)
Inspect and extract archives
List an archive without extracting it:
tar -tf <archive.tar> | head
Extract it into a chosen directory:
mkdir -p <destination>
tar -xf <archive.tar> -C <destination>
Keep the original directory hierarchy when extracting archives, because annotations and SoccerNet development tools commonly identify games through their relative paths.
List a ZIP archive without extracting it:
unzip -l <archive.zip> | head
Extract it into a chosen directory:
unzip <archive.zip> -d <destination>
Use tar for frames-720p-2fps and clips-720p-10s. Use unzip for clips-720p-10s-zip.
Annotations and SoccerNet tools
This repository focuses on the protected video media and derived visual representations. SoccerNet annotations, split definitions, development kits, and task-specific utilities are maintained through the broader SoccerNet resources:
- SoccerNet website
- SoccerNet Python package and development repository
- SoccerNet organization on Hugging Face
Install the SoccerNet Python package with:
python -m pip install --upgrade SoccerNet
The standard game splits can be queried with:
from SoccerNet.utils import getListGames
train_games = getListGames(split="train")
valid_games = getListGames(split="valid")
test_games = getListGames(split="test")
challenge_games = getListGames(split="challenge")
NDA, copyright, and user responsibility
The videos in this repository are broadcast media and may be protected by copyrights and other rights belonging to broadcasters, leagues, clubs, production companies, event organizers, or other third parties.
The SoccerNet NDA provides a controlled contractual mechanism for sharing the dataset with approved researchers who agree to confidentiality, non-redistribution, research-only, and non-commercial-use obligations. The NDA does not circumvent, waive, transfer, sublicense, or grant any copyright or other intellectual-property right in the videos, audio, graphics, trademarks, personal likenesses, or other protected material contained in the dataset.
In particular, access to this repository does not by itself authorize:
- redistribution of the dataset or any substantial portion of it;
- publication of source videos, clips, frames, or compilations containing protected footage;
- public screening, broadcasting, or commercial exploitation;
- removal of watermarks, broadcaster graphics, or rights-management information;
- use that infringes copyright, privacy, publicity, database, contractual, or other applicable rights; or
- any activity prohibited by the NDA or by applicable law.
Each user is solely responsible for determining whether their acquisition, storage, processing, publication, presentation, and other use of the dataset is lawful in every jurisdiction applicable to them. Users must obtain any additional permissions, licenses, ethics approvals, institutional approvals, or legal advice required for their intended use.
To the fullest extent permitted by applicable law, the SoccerNet team, its contributors, and affiliated institutions cannot be held responsible or liable for use of the dataset that violates the NDA, third-party rights, institutional rules, or any national, regional, or local law applicable to the user. Access may be suspended or revoked if the access conditions are breached.
If this dataset card conflicts with the executed NDA, the executed NDA controls.
Permitted use summary
Subject to the executed NDA and applicable law, access is intended for:
- academic and scientific research;
- non-commercial experimentation and benchmarking;
- development and evaluation of methods for sports video understanding; and
- publication of research results that do not redistribute or expose protected media beyond what is legally permitted.
This summary is informational and does not replace the NDA.
Security and responsible handling
- Store the data on access-controlled systems.
- Limit access to the approved Hugging Face account holder.
- Do not place the data in public buckets, repositories, shared drives, or unrestricted servers.
- Do not commit Hugging Face tokens or downloaded media to source-control repositories.
- Remove local and cached copies when they are no longer required or when access is revoked.
- Contact the SoccerNet team promptly if protected files are accidentally exposed.
Citation
If you use these videos or their derived representations in academic work, cite the relevant SoccerNet publications.
@inproceedings{giancola2018soccernet,
title = {SoccerNet: A Scalable Dataset for Action Spotting in Soccer Videos},
author = {Giancola, Silvio and Amine, Mohieddine and Dghaily, Tarek and Ghanem, Bernard},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
year = {2018}
}
@inproceedings{deliege2021soccernetv2,
title = {SoccerNet-v2: A Dataset and Benchmarks for Holistic Understanding of Broadcast Soccer Videos},
author = {Deli\`ege, Adrien and Cioppa, Anthony and Giancola, Silvio and Seikavandi, Meisam J. and Dueholm, Jacob V. and Nasrollahi, Kamal and Ghanem, Bernard and Moeslund, Thomas B. and Van Droogenbroeck, Marc},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
year = {2021}
}
Contact
For questions about access, the NDA, or an existing request, contact:
For documentation, benchmarks, and community resources, visit soccer-net.org.
- Downloads last month
- 8