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.

SoccerNet-GAR: Pixels or Positions? Benchmarking Modalities in Group Activity Recognition

SoccerNet-GAR is a large-scale multimodal dataset for Group Activity Recognition (GAR) built from all 64 matches of the FIFA World Cup 2022 tournament. It provides synchronized broadcast video and player tracking data for 87,939 annotated group activities across 10 action classes, enabling direct comparison between video-based and tracking-based approaches.

Dataset Details

Description

SoccerNet-GAR is the first dataset to provide synchronized tracking and video modalities for the same action instances in group activity recognition. For each annotated event, a 4.5-second temporal window is extracted from both the broadcast video and player tracking streams, centered on the event timestamp. In the original sampled representations (frames and tracking), 16 samples are taken with a 9-frame interval (approximately 3.3 samples per second over a 4.5-second first-to-last-sample span). The additional videos and tracking-full representations retain the full 144-frame native-rate window; see Native-FPS MP4 clips below.

The dataset contains two input modalities:

Video Modality: Broadcast footage at 720p resolution. Each frame is part of a temporal sequence sampled within the event window, capturing appearance cues, scene context, and visual motion patterns.
Tracking Modality: 2D player positions and 3D ball coordinates sampled at 30 fps, automatically extracted from broadcast footage and manually refined by annotators. Player positions span x in [-60, 60]m, y in [-42, 41]m; ball positions include height z in [-8, 25]m. Each entity state encodes spatial coordinates, entity identity (one-hot encoding), and motion dynamics (displacement vectors between consecutive frames). Positional role metadata (goalkeeper, defender, midfielder, forward) is provided for each player.

Property	Value
Curated by	KAUST, University of Liege
Original Data Source	Gradient Sports (formerly PFF FC)
Total Events	87,939
Matches	64 (Football World Cup 2022)
Action Classes	10
Modalities	Video + Tracking
Avg. Events per Match	1,374

Sources

Repository: https://github.com/drishyakarki/pixels_vs_positions
Paper: Pixels or Positions? Benchmarking Modalities in Group Activity Recognition (arXiv:2511.12606)
OpenSportsLib: https://github.com/OpenSportsLab/opensportslib

How to Use

The recommended way to download and use SoccerNet-GAR is through OpenSportsLib.

The two modalities, video and tracking, are available in four OpenSportsLib-ready representations:

Branch	Representation	Temporal sampling	OSL input type
`frames`	Original sampled RGB frame arrays (`.npy`, `16 × 224 × 224 × 3`, `uint8`)	16 frames, stride 9	`frames_npy`
`tracking`	Original sampled player/ball tracking (`.parquet`)	16 sampled frame slots	`tracking_parquet`
`videos`	Native-resolution MP4 action clips	144 consecutive source frames at native timing, approximately 29.97 FPS	`video`
`tracking-full`	Full-rate player/ball tracking (`.parquet`)	144 native frame slots, approximately 29.97 FPS	`tracking_parquet`

All four representations retain the same 87,939 action identities, split assignments and labels. Match samples across branches by sample ID within each split. Use frames and tracking for the original sampled benchmark inputs; use videos and tracking-full when you need the intervening native-rate frames.

The original data representations remain available through the frames-raw and tracking-raw branches.

Installation

pip install opensportslib

Video (Pixels)

Download the video modality from the canonical frames branch:

from opensportslib.tools import download_dataset_split_from_hf

download_dataset_split_from_hf(
    repo_id="OpenSportsLab/SoccerNet-GAR",
    revision="frames",
    split="train",
    output_dir="SoccerNet-GAR",
)

Replace train with valid or test to download the other splits.

Train and evaluate:

from opensportslib import model

if __name__ == "__main__":
    myModel = model.classification(
        config="path/to/sngar-frames.yaml",
        data_dir="SoccerNet-GAR",
    )

    myModel.train(
        train_set="SoccerNet-GAR/annotations_train.json",
        valid_set="SoccerNet-GAR/annotations_valid.json",
        use_ddp=False,
    )

    myModel.infer(
        test_set="SoccerNet-GAR/annotations_test.json",
    )

Tracking (Positions)

Download the tracking modality from the canonical tracking branch:

from opensportslib.tools import download_dataset_split_from_hf

download_dataset_split_from_hf(
    repo_id="OpenSportsLab/SoccerNet-GAR",
    revision="tracking",
    split="train",
    output_dir="SoccerNet-GAR",
)

Replace train with valid or test to download the other splits.

Train and evaluate:

from opensportslib import model

myModel = model.classification(
    config="path/to/classification_tracking.yaml",
    data_dir="SoccerNet-GAR",
)

myModel.train(
    train_set="SoccerNet-GAR/annotations_train.json",
    valid_set="SoccerNet-GAR/annotations_valid.json",
)

myModel.infer(
    test_set="SoccerNet-GAR/annotations_test.json",
)

Configuration files for reproducing the experiments in the paper are available in the pixels_vs_positions repository.

Native-FPS MP4 clips (`videos`)

The videos branch contains one MP4 clip per classification action, regenerated from the source videos in SNGAR-Action-Spotting. Clips preserve the source resolution, native frame timing and all intervening frames. They are encoded with H.264/libx264, CRF 18 and yuv420p, without audio. The source timestamps determine timing; clips are not forced to 30 FPS.

Each clip uses the full sampling window: from the first of the 16 sampled frames through one sampling interval after the last, exclusive. With stride 9, this contains 144 native frames, approximately 4.805 seconds at 29.97 FPS. The approximately 4.5-second span between the first and last sampled frames in the original representation is therefore shorter than the full native window.

The MP4s preserve action correspondence and validated temporal alignment. They do not claim pixel-identical reproduction of the original 224 × 224 NumPy arrays; resizing, historical preprocessing and lossy MP4 encoding can differ.

from opensportslib.tools import download_dataset_split_from_hf

video_files = download_dataset_split_from_hf(
    repo_id="OpenSportsLab/SoccerNet-GAR",
    revision="videos",
    split="test",
    output_dir="SoccerNet-GAR",
    download_format="parquet",
)

Full-FPS tracking (`tracking-full`)

The tracking-full branch provides one Parquet clip per classification action, extracted from the original spotting tracking tables. Each clip contains 144 native frame slots over the corresponding full action window, including the frames omitted by the stride-9 sampled representation. Observed source values and columns are preserved; frameNum and videoTimeMs retain the source frame numbers and absolute timestamps.

Some source windows contain missing tracking observations. The release preserves these slots and adds the boolean tracking_valid column:

true: an observed source tracking row.
false: a missing source row, represented by a known frameNum, null videoTimeMs, period = -1 and empty player/ball arrays.

Missing positions are not interpolated. Use the validity flag when constructing masks or computing motion features; a native frame slot is not necessarily an observed tracking row. A valid row can still have missing individual entities. The OSL input type remains tracking_parquet; tracking-full is the branch name.

from opensportslib.tools import download_dataset_split_from_hf

tracking_files = download_dataset_split_from_hf(
    repo_id="OpenSportsLab/SoccerNet-GAR",
    revision="tracking-full",
    split="test",
    output_dir="SoccerNet-GAR",
    download_format="parquet",
)

For both examples, replace test with train or valid to download other splits. Use the returned download information to locate the reconstructed OSL annotation manifest. Configure temporal sampling explicitly: frame_interval: 1 retains consecutive native frames; stride 9 samples the full-rate inputs back down. Changing only the branch does not automatically change a model's temporal input.

Shards and Parquet metadata

Both new branches use OSL's Parquet + WebDataset distribution format. Each split (train: 62,159, valid: 12,091, test: 13,689 samples) has:

<split>/
  metadata.parquet
  shard_manifest.parquet
  shards/
    *.tar

The TAR shards contain the MP4 or per-action tracking Parquet files together with OSL JSON sidecars. The split metadata and shard manifest support sample lookup and OSL downloads. Sample metadata retains action labels and conversion provenance, including source hashes/revisions where available, native frame windows, timing and conversion identity. Original source paths in provenance are informational, not paths that consumers need to access.

The published benchmark results below refer to the original sampled representations; they are not new evaluations of the full-rate branches.

Dataset Structure

Action Classes

The dataset contains 10 action classes reflecting common football events:

Class	Count	Proportion
PASS	57,521	65.4%
TACKLE	10,943	12.4%
OUT	5,873	6.7%
HEADER	5,723	6.5%
THROW IN	2,598	3.0%
CROSS	2,175	2.5%
FREE KICK	1,788	2.0%
SHOT	1,041	1.2%
GOAL	30	0.2%
HIGH PASS	12	0.1%

The dataset exhibits severe class imbalance (646:1 ratio between PASS and HIGH PASS), reflecting the natural distribution of football events.

Splits

Data is split at the match level to prevent leakage:

Split	Matches	Events	Proportion
Train	45	62,159	70.7%
Validation	9	12,091	13.7%
Test	10	13,689	15.6%

Data Quality

Player tracking completeness: 99.9% of 1,485,008 frames contain all 11 players per team.
Ball visibility: 93.4% of frames contain ball tracking data.
Event-level ball coverage: 85.9% of annotated events have complete ball tracking within their temporal window.

Branches

This repository is organized into the following branches:

Branch	Contents
`main`	Dataset card and documentation.
`frames`	Original sampled video frames: 16-frame NumPy arrays per event.
`tracking`	Original sampled tracking: 16-frame Parquet clips per event.
`videos`	Native-resolution, native-FPS MP4 clips: 144 consecutive frames per event.
`tracking-full`	Full-FPS tracking Parquet clips: 144 native frame slots with `tracking_valid` masks.
`frames-raw`	Original/raw video representation, with one NPY clip per event.
`tracking-raw`	Original/raw tracking representation.

Use frames and tracking for the original sampled benchmark representations, or videos and tracking-full for native-rate action windows. The *-raw branches provide access to the original raw representations.

Benchmark Results

Pixels vs. Positions

Modality	Model	Params	Bal. Acc.	F1	Training
Tracking	GIN + MaxPool + Positional Edges	180K	77.8%	57.0%	4 GPU hours
Video	VideoMAEv2-B (finetuned)	86.3M	60.9%	50.1%	28 GPU hours

The tracking model outperforms the video baseline by 16.9 percentage points in balanced accuracy and 6.9 percentage points in macro F1 while using 479x fewer parameters and training 7x faster.

Per-Class Comparison (Test Set, Balanced Accuracy)

Class	Samples	Tracking	Video
PASS	9,009	81.1	77.6
TACKLE	1,690	54.0	32.2
OUT	884	94.2	75.8
HEADER	867	65.2	66.3
THROW IN	12	84.2	78.6
CROSS	392	86.7	77.2
FREE KICK	347	90.4	79.4
SHOT	272	76.3	63.4
GOAL	186	73.3	16.7
HIGH PASS	30	83.3	41.7

Tracking dominates on 9 of 10 classes, with its largest gains on less frequent classes like GOAL (+56.7 pp) and HIGH PASS (+41.7 pp). Video shows a slight advantage only on HEADER (+1.1 pp). Tracking models learn discriminative features even in severely data-scarce regimes (GOAL: 73.3%, HIGH PASS: 83.3%), whereas video models collapse on these classes (16.7% and 41.7%).

Uses

Direct Use

Benchmarking video-based vs. tracking-based group activity recognition
Training and evaluating GAR models on football broadcast data
Studying multimodal fusion approaches combining visual and positional features
Analyzing spatial interaction patterns in team sports

Dataset Creation

Curation Rationale

No standardized benchmark previously existed that aligns broadcast video and tracking data for the same group activities. This made fair, apples-to-apples comparison between video-based and tracking-based approaches impossible. SoccerNet-GAR was created to fill this gap by providing synchronized multimodal observations under a unified evaluation protocol.

Source Data

The dataset was constructed from the PFF FC website (now Gradient Sports), which provides broadcast videos, player tracking data, and event annotations across all 64 FIFA World Cup 2022 tournament matches.

Data Cleaning and Alignment

Event annotations are aligned with both input modalities by merging them with tracking streams using UTC timestamps. Three successive filters ensure data quality:

Temporal alignment: Events where no tracking frame falls within a 10 ms tolerance of the event timestamp are removed.
Modality coverage: Events lacking corresponding data in either modality are discarded.
Duplicate resolution: When a single timestamp is annotated with more than one action class (e.g., a goal also labeled as a shot), only the most semantically specific label is retained based on a predefined priority ordering.

Together, these filters remove 6,346 events (6.8% of raw annotations), yielding the final dataset of 87,939 annotated group activities.

Annotation Process

Event annotations with precise timestamps were created by trained annotators and verified through quality control procedures by PFF FC using both video and tracking views. Each event is labeled with one of 10 group activities and temporally marked at the moment of occurrence.

Comparison with Existing Datasets

Dataset	Year	Domain	Events	Classes	Modalities
CAD	2009	Pedestrian	2,511	5	V
Volleyball	2016	Volleyball	4,830	8	V
SoccerNet	2018	Football	6,637	3	V
NBA	2020	Basketball	9,172	9	V
SoccerNet-v2	2021	Football	110,458	17	V
NETS	2022	Basketball	61,053	3	T
SoccerNet-BAS	2024	Football	11,041	12	V
Cafe	2024	Indoor	10,297	6	V
FIFAWC	2024	Football	5,196	12	V
SoccerNet-GAR	2026	Football	87,939	10	V + T

SoccerNet-GAR is the second largest GAR dataset (after SoccerNet-v2) and the only one providing synchronized video and tracking modalities for the same action instances.

Citation

@article{karki2025pixels,
  title={Pixels or Positions? Benchmarking Modalities in Group Activity Recognition},
  author={Karki, Drishya and Ramazanova, Merey and Cioppa, Anthony and Giancola, Silvio and Ghanem, Bernard},
  journal={arXiv preprint arXiv:2511.12606},
  year={2025}
}

Authors

Drishya Karki (KAUST)
Merey Ramazanova (KAUST)
Anthony Cioppa (University of Liege)
Silvio Giancola (KAUST)
Bernard Ghanem (KAUST)

Contact

SoccerNet Challenge 2027 (GAR). The baseline uses the 16 sampled time steps from the tracking branch. Participants are welcome to use any available representation, including the full-rate videos and tracking-full branches, or to combine video and tracking inputs. The challenge uses the existing test split. OSL-format manifests for train, valid, and test across all four representations are available in the SN-GAR-2027 repository; see RULES.md for rules and submission details.

Downloads last month: 729

Total file size:

21.9 kB

Collection including OpenSportsLab/SoccerNet-GAR

OSL-ready datasets

Collection

List of OpenSportsLib ready datasets • 5 items • Updated about 1 month ago

Paper for OpenSportsLab/SoccerNet-GAR

Pixels or Positions? Benchmarking Modalities in Group Activity Recognition

Paper • 2511.12606 • Published Nov 16, 2025

Request access to SoccerNet GAR dataset