OpenSportsLab/soccernetpro-classification-GAR

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. Within each window, 16 samples are taken at 30 fps with a 9-frame interval (effectively sampled at approximately 3.3 fps over 4.5 seconds).

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)

How to Use

Results from the paper can be reproduced using OpenSportsLib. Configuration files for all experiments are provided in the pixels_vs_positions repository.

Tracking (Positions)

Download data from the tracking-parquet branch:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="OpenSportsLab/soccernetpro-classification-GAR",
    repo_type="dataset",
    revision="tracking-parquet",
    local_dir="sngar-tracking",
)

Extract the splits:

unzip train.zip && unzip valid.zip && unzip test.zip
rm train.zip valid.zip test.zip

Train and evaluate:

from opensportslib import model

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

myModel.train(
    train_set="sngar-tracking/annotations_train.json",
    valid_set="sngar-tracking/annotations_valid.json",
)

myModel.infer(test_set="sngar-tracking/annotations_test.json")

Video (Pixels)

Download data from the frames branch:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="OpenSportsLab/soccernetpro-classification-GAR",
    repo_type="dataset",
    revision="frames",
    local_dir="sngar-frames",
)

The training split is stored as a multi-part zip. Reassemble and extract:

cat train.zip.part_aa train.zip.part_ab > train.zip
unzip train.zip && unzip valid.zip && unzip test.zip
rm train.zip.part_aa train.zip.part_ab train.zip valid.zip test.zip

Train and evaluate:

from opensportslib import model

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

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

    myModel.infer(test_set="sngar-frames/annotations_test.json")

See the pixels_vs_positions repository for the specific config files needed to reproduce each experiment in the paper.

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	188	0.2%
HIGH PASS	89	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.
paper-data	The exact dataset needed to reproduce the results in the paper. Contains broadcast videos (1 npy clip per event) and tracking files (1 parquet file per full match).
frames	1 npy clip per event for the video modality. Annotations are in SoccerNetPro format.
tracking-parquet	1 parquet clip per event for the tracking modality. Annotations are in SoccerNetPro format.
multimodal-data	Combined video (npy) and tracking (parquet) data with 1 file per event per modality. Uses a unified annotation file for both modalities in SoccerNetPro format.

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:

1. Temporal alignment: Events where no tracking frame falls within a 10 ms tolerance of the event timestamp are removed.
2. Modality coverage: Events lacking corresponding data in either modality are discarded.
3. 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

• drishya.karki@kaust.edu.sa / karkidrishya1@gmail.com
• silvio.giancola@kaust.edu.sa