Please submit your questions to abc2027@autocare.ai or trackid3x3-challenge-2026@g.sp.m.is.nagoya-u.ac.jp with the subject line "ABC Challenge."
Please email the organizers if there are any changes to team members, email addresses, or other details.
TrackID3x3 Challenge FAQ
Updated: November 2, 2026
These FAQs consolidate the current challenge rules and clarifications. They apply equally to all teams and take precedence over older wording on the same topics.
Important update (November 2): All teams must submit a paper to be eligible for results evaluation. This supersedes earlier TrackID3x3 guidance limiting mandatory paper submission to the top five teams. Reproducibility materials remain mandatory for the provisional top five teams and optional for other teams. See Questions 13–14 for deadlines and submission instructions.
1. What data may be used for training and fine-tuning?
Use the official Training set, including ordinary augmentation of that set. Direct additional training on external real-world footage or annotations, or ordinary augmentations of those external data, is not permitted. The exceptions for public pretrained checkpoints and eligible synthetic data are explained below.
2. Are publicly available pretrained models allowed?
Yes. Publicly available pretrained checkpoints may be used regardless of their original application domain, provided they were not trained or fine-tuned on the Challenge sequences or their corresponding annotations. Any further training must follow the permitted-training-data rules. Disclose the exact checkpoints, their sources, and the data used for additional training.
3. Is synthetic training data allowed?
Yes. Synthetic data may be generated independently, from the official Training set, or using public source material permitted for this use. Non-public real-world footage or annotations, Challenge evaluation sequences, synchronized views, and corresponding annotations must not be used as sources, directly or indirectly. This exception does not allow direct training on external real footage or its ordinary augmentations. Disclose the sources and generation/training procedures. These conditions also apply to synthetic data created before the policy update.
4. What processing of the Challenge videos is permitted?
Both online and offline inference are permitted, including automatic processing of a complete Challenge video. Strictly causal or real-time processing is not required. Normal inference-time tracking-state updates are allowed. Self-training, training on pseudo-labels, and model-weight fine-tuning using Challenge videos are prohibited.
5. What manual input is permitted?
Manual selection of fixed court landmarks for camera/court calibration is permitted and must be described in the method report. Manual annotation, identity assignment, or correction of players or their trajectories in the Challenge videos is prohibited. Submitted player predictions must be generated by the reported automatic method.
6. May external annotations corresponding to the Challenge videos be used?
No. Corresponding annotations, synchronized-view ground truth, or existing ground-truth trajectories must not be used for training, pseudo-label generation, or generating submitted predictions, even if accessible from another source.
7. Which court-registration file should be used?
Use the latest official court-registration file. A corrected version was released on October 21. Teams may also refine calibration using visible fixed court landmarks, subject to the manual-input rule above. A correction to court-registration metadata does not change the ground-truth construction described below.
8. How is the challenge evaluated?
Ranking is based on TI-HOTA using the public evaluation implementation and a 100 cm court-distance parameter. TI-AssA is averaged equally over ground-truth player identities, rather than weighted by the number of matched instances; identities with no successful matches contribute zero. Player identity is defined by team affiliation and jersey number. All frames are evaluated, including frames before play begins. Preserve the original frame numbering.
Evaluation implementation: https://github.com/open-starlab/TrackID3x3/blob/main/script/compute_TI-HOTA.ipynb
9. What do the reference court positions represent?
The Challenge reference positions and identities are derived from time-aligned top-view annotations, using the same coordinate system and player-position convention as the official Training set. The position reference is the bottom-center of the annotated player bounding box in the top-view footage. It is not independently reconstructed from bounding boxes in the oblique-view Challenge videos.
10. Are there hardware or runtime restrictions?
No specific hardware restriction, runtime limit, or real-time requirement is imposed. Runtime is not a ranking or eligibility criterion, and a runtime report is not mandatory. Hardware and approximate end-to-end runtime may be reported for reproducibility; please state which processing steps are included.
11. How and when should predictions be submitted?
Submit one ZIP file containing IMG_0114.txt and IMG_0115.txt directly, with no additional folders, through the official form. Each TXT file must be comma-separated and headerless, with these five fields in order:
frame_id,court_x,court_y,team_id,jersey_number
Frame IDs start at 1. Court coordinates are in centimeters and follow the Training ground-truth coordinate system. Team IDs are A/B; no separate track-ID column is required. Use the registered team name and contact email.
Results form: https://forms.gle/rMwK4pHm1Xp2joV27
Deadline: November 3, 2026, 23:59 AoE (UTC−12).
The current upload limit is 10 MB. Please email the organizers if you encounter an upload problem.
12. Can a submission be replaced, and is intermediate feedback available?
Yes. Multiple submissions are permitted before the deadline; only the latest submission received before the deadline is evaluated. No intermediate scores or ranking feedback are provided. The challenge uses an official Training set and a Challenge evaluation set, with no separate intermediate Test-set leaderboard.
13. Who must submit a paper and reproducibility materials?
All teams must submit a full challenge paper by November 10, 2026, 23:59 AoE (UTC−12), to be eligible for results evaluation, regardless of their ranking. Paper quality and acceptance do not affect the numerical ranking, which is determined by TI-HOTA. Papers are reviewed through the ABC paper-review process.
Separately, reproducibility materials are mandatory for the provisional top five teams and optional for other teams. Required materials include code, model checkpoints, configuration, and a README sufficient to reproduce the submitted predictions. Submit these materials through the CMT supplementary-file option by November 10, 2026, 23:59 AoE (UTC−12). Preserve the exact versions used to generate the predictions submitted by November 3. Large files may be provided through fixed-version download links kept accessible during the checks. Public code release is not required.
Awards remain subject to rule compliance and successful integrity/reproducibility checks. The requirement to submit a paper is separate from paper acceptance.
14. Is a separate CMT submission required, and what are the deadlines?
Yes. Completing the Team Registration Google Form does not replace the CMT paper submission. Every team must create a submission with the title of its work in ABC2027 CMT and select “Challenge Track” by November 3, 2026, 23:59 AoE (UTC−12). The full paper must be uploaded by November 10, 2026, 23:59 AoE (UTC−12). Supplementary files can also be uploaded or updated until the November 10 deadline.
CMT: https://cmt3.research.microsoft.com/IABC2027/Submission/Index
Prediction files must be submitted separately through the Results Submission Google Form by November 3, 2026, 23:59 AoE (UTC−12), as explained in Question 11. Creating a CMT account alone does not constitute a paper submission.
For a missing team-registration confirmation or a correction to an existing registration, email the organizers with the registered team name and contact email.
15. How should participants contact the organizers?
For this edition, official participant communication will take place by email. We will not open a separate Discord, Slack, or other community chat channel. Please send questions to trackid3x3-challenge-2026@g.sp.m.is.nagoya-u.ac.jp. General clarifications will be shared with all teams in a form that does not identify the questioner or disclose team-specific methods.
16. Where will accepted challenge papers be published, and what fees apply?
Accepted challenge papers will be published in IJABC (J-STAGE). There is no IEEE publication option for this edition of the challenge.
Conference registration fees are required once a paper is accepted. No additional IJABC publication fee applies beyond conference registration. Please consult the official ABC2027 conference registration information for the applicable rates and registration arrangements.