2025 Challenge Track Description
Track 1: Multi-Camera 3D Perception
Challenge Track 1 involves synthetic animated people data in multiple indoor settings generated using the NVIDIA Omniverse Platform. The 2025 edition expands last year’s corpus with:
| Split | Hours | Cameras | Scenes | Resolution / FPS | Objects in Training GT* | File Size |
| Train + Val | ≈42 h | 504 | 19 indoor layouts (warehouse, hospital, retail, office) | 1080 p @ 30 fps | 363 instances (292 person + service robots / forklifts) | 74 GB ( depth maps optional, >3 TB) |
* Counts refer to the training + validation ground-truth only; the test split is hidden before the release of evaluation system.
Each scene provides temporally-synchronized RGB video, camera calibration, a top-down map, and per-frame 2D/3D annotations. Depth maps (PNG-in-HDF5) are included but very large; feel free to ignore them if storage or I/O is a concern.
- Task
Teams should detect every object and keep the same identity ID while they move within and across all cameras in a scene.
- Submission Format
For compatibility with the official evaluation server, results must be a single plain-text file (track1.txt) where each line describes one detection:
〈scene_id〉 〈class_id〉 〈object_id〉 〈frame_id〉 〈x〉 〈y〉 〈z〉 〈width〉 〈length〉 〈height〉 〈yaw〉
| Field | Type | Description |
| scene_id | int | Unique identifier for each multi-camera sequence. |
| class_id | int | Starting from zero, denoting an object’s category. (Person→0, Forklift→1, NovaCarter→2, Transporter→3, FourierGR1T2→4, AgilityDigit→5.) |
| object_id | int | Positive, unique ID per scene & class. Remains constant across all cameras within the same scene and class. |
| frame_id | int | Zero-based frame index within that scene. |
| x, y, z, | float | 3D coordinates of the bounding-box centroid in the world coordinate system which is in meters. |
| width, length, height | float | Box dimensions in meters along its x (width), y (length) and z (height) axes of the object-centered coordinate system, with the origin at the centroid. |
| yaw | float | Euler angle in radians about the y-axis of the object-centered coordinate system defining the box’s heading in the world coordinate system. (Pitch and roll are assumed zero.) |
Example: in scene 0, if a Person is assigned obj_id = 5, then a Forklift cannot use obj_id = 5 (it must use a different ID, e.g. 6).
Archive the text file as track1.zip or track1.tar.gz before uploading.
- Evaluation
Scores are computed with 3D HOTA [1], which jointly balances detection, association and localization quality. HOTA score will be computed per class within a scene which will be averaged. A weighted average will then be computed on these scores across all scenes based on the total no. of objects. 3D IoU will be used for matching GT & prediction objects.
- Leaderboard = raw HOTA on the hidden test set.
- Online-tracker bonus: If your paper + code prove that only past frames are used, a +10 % multiplicative bonus is applied when deciding the final winner and runner-up (the public leaderboard itself shows the un-bonused score).
Example: Team A (offline) = 69 % HOTA; Team B (online) = 61 % ⇒ bonus → 67.1 % HOTA. Team B ranks higher in the final award list.
- Data Access
Note: Depth files are huge (> 3 TB). If bandwidth or disk is limited, download only the other files.
By downloading you agree to the Physical AI Smart Spaces licence (CC-BY 4.0).
References
[1] J. Luiten et al., “HOTA: A Higher Order Metric for Evaluating Multi-Object Tracking,” IJCV, 2021.