📢 Group Activity Recognition (GAR) Challenge! 🚀
📅 Submission deadline: April 25, 2027, 23:59 Anywhere on Earth (AoE; UTC−12).
👥 Task lead: Silvio Giancola
🏠 Challenge website · 🗂️ Data on Hugging Face · 💻 Baseline · 📚 Rules · 📊 Evaluation server · 🧮 Evaluation code
Predict one of ten group activity classes for each soccer clip using video pixels, player positions, or both. All modalities share one leaderboard ranked by balanced accuracy. The supplied baseline uses 16 sampled frames of player and ball positions from the tracking branch; participants are welcome to explore the videos and tracking-full branches or combine available representations. The task and reference method are described in Pixels or Positions? Benchmarking Modalities in Group Activity Recognition.
The gated SoccerNet-GAR dataset provides train, valid, and test material. Request access on Hugging Face and choose among the tracking, frames, videos, and tracking-full branches. The OpenSportsLib JSON manifests for train, validation, and test include ground-truth labels and references to all four modality inputs; the media files are not copied into this repository. Approved users can already access the labels on Hugging Face, so the same test labels are published here to support reproducible evaluation. Follow the no-test-training rule.
The reference method is the GIN + MaxPool + positional-edge tracking model from Karki et al.. The baseline guide provides its trained checkpoint, test predictions, submission ZIP, and instructions to reproduce inference with OpenSportsLib. The supplied predictions score 78.64% balanced accuracy on the challenge test set.
Submit predictions through the CodaBench evaluation server. The evaluation guide documents the prediction format and scorer. The scorer calls OpenSportsLib's classification metrics and reports balanced accuracy, macro F1, accuracy, and ten per-class recalls as percentages; balanced accuracy remains the ranking metric.
Please cite the GAR benchmark paper in its CVPR Workshops version:
@InProceedings{Karki_2026_CVPR,
author = {Karki, Drishya and Ramazanova, Merey and Cioppa, Anthony and Giancola, Silvio and Ghanem, Bernard},
title = {Pixels or Positions? Benchmarking Modalities in Group Activity Recognition},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2026},
pages = {9998--10008},
url = {https://openaccess.thecvf.com/content/CVPR2026W/CVsports/html/Karki_Pixels_or_Positions_Benchmarking_Modalities_in_Group_Activity_Recognition_CVPRW_2026_paper.html}
}
Follow the SoccerNet challenge website and Discord for challenge announcements and discussion.