The challenge dataset is organized into three main splits:
dataset/
│
├── training/
│ ├── top_view_videos/
│ ├── player_annotations/
│ ├── court_registration/
│ └── game_rosters/
│
├── test/
│ ├── oblique_view_videos/
│ ├── court_registration/
│ └── game_rosters/
│
└── final_challenge/
├── oblique_view_videos/
├── court_registration/
└── game_rosters/
This is a website-friendly representation of the logical dataset structure; the proposal describes these resources by split but does not specify these exact folder names.
Structure Description
The Training Set contains published top-view basketball videos together with full player annotations, court-registration information, and per-game rosters. These data can be used for model training and local validation.
The Test Set contains two unseen oblique-view videos together with court-registration metadata and player rosters. Ground-truth player annotations remain hidden and the split is intended for development evaluation.
The Final Challenge Set contains two additional oblique-view videos and is used for official challenge ranking. Ground-truth annotations are not released to participants.
For every video, the roster defines the six valid player identities. Each player's persistent identity is represented by the combination of team affiliation and jersey number.
Court-registration metadata is provided so participants can project detected players from image coordinates onto a standardized two-dimensional basketball court.
Published papers about the dataset
1. K. Yamada, L. Yin, Q. Hu, N. Ding, S. Iwashita, J. Ichikawa, K. Kotani, C. Yeung, and K. Fujii “TrackID3x3: A Dataset and Algorithm for Multi-Player Tracking with Identification and Pose Estimation in 3x3 Basketball Full-court Videos.” Proceedings of the 8th International ACM Workshop on Multimedia Content Analysis in Sports (MMSports '25), pp. 163–173, 2025. DOI: 10.1145/3728423.3759400
This publication introduces the TrackID3x3 dataset, fixed-camera Track-ID task, baseline tracking and identification pipeline, and TI-HOTA evaluation metric. Challenge participants using the dataset are required to cite this publication.