Datasets:
🔍 Are These Cameras in the Same Universe? *(Not a Multiverse… right?)*
Dear organizers and participants,
During an EDA on the Warehouse_003 scene in the MTMC_Tracking_2025 dataset, I met a curious phenomenon that may indicate either a simulation inconsistency or a misalignment in camera synchronization.
Specifically, I was comparing object trajectories across Camera_0014 and Camera_0039. At first (zero-frame), everything seemed in order — the initial object placements and identities were consistent. However, as time progresses, the motion patterns begin to diverge significantly. What initially appeared to be the same scene captured from different angles slowly transforms into two distinct evolutions of object behavior.
My initial assumption was a frame synchronization issue. However, after closer inspection, the trajectories themselves — particularly of a large man in pink wearing headphones pushing a cart — follow entirely different paths in the two camera views, despite starting from identical initial positions. This divergence appears to go beyond temporal misalignment and suggests a fundamental desynchronization in the underlying simulation.
(🎥 I’ve attached a GIF to visually demonstrate this discrepancy — the subject in question follows notably different trajectories in each view.)
This raises a few questions:
- ❓ Are these two camera views (
Camera_0014andCamera_0039) — and potentially other camera pairs in the same scene — intended to be synchronized representations of the same physical space? - 🔁 Could this be due to simulation desynchronization during the synthetic data generation process?
- 🗺️ Is there any metadata (e.g., camera grouping, synchronization domains) that would clarify this behavior?
As a possible explanation, I wonder whether this may be related to non-determinism in Omniverse Replicator. Even when using the same random seed, some aspects of the simulation (e.g., physics, agent policies, rendering order) may produce slightly different results across runs, especially if the simulation for each camera was rendered separately. If that’s the case, this might explain the gradual divergence between seemingly identical scenes.
If I’ve missed anything in the documentation or configuration that could clarify this, I would greatly appreciate any insights.
Thank you again for providing this dataset — it’s a valuable resource for the multi-camera tracking research community.
Best regards,
visionNoob
Dear @visionNoob ,
Thank you for your detailed analysis and for bringing this to our attention — and apologies for the delayed response.
We’re currently pushing updated train/val annotations to Hugging Face and removing misaligned video sequences (and depth maps) to address this issue. Please refer to the changelog on the Hugging Face repository for details once the update is live.
In light of this, the total submission limit of Track 1 for the 2025 AI City Challenge will be increased to help teams adjust and re-evaluate their models. We’ve confirmed that the ground truth of test set remains unaffected.
For future inquiries, please feel free to also email us at [email protected] to ensure we can respond promptly.
Thanks,
Thomas
@zhengthomastang
for training set, my analysis is below
❌ 000 -> total 51 -> up to 25 (0000-0025)
✅ 001 -> total 42 -> fine
✅ 002 -> total 46 -> fine
❌ 003 -> total 45 -> up to 30 (0000-0029)
✅ 004 -> total 42 -> fine!!
❌ 005 -> total 48 -> up to 30 (0000-0029)
✅ 006 -> total 10 -> fine
✅ 007 -> total 12 -> fine
✅ 008 -> total 9 -> fine
✅ 009 -> total 11 -> fine
❌ 010 -> total 40, up to 25 (0000-0024)
❌ 011 -> total 46, up to 16 (0000-0015)
✅ 012 -> total 12 -> fine
✅ 013-> total 12 -> fine
✅ 014 -> total 12 -> fine
@visionNoob
Following up on the earlier message — all train/val scenes of MTMC_Tracking_2025 have now been fully updated. This includes corrected annotations, aligned calibration files, and the removal of misaligned video and depth map sequences.
@zhengthomastang
Yes i understood, by the way.
Depite of "fully updating", I'm just wordering that remained 50 cameras of "Warehouse_000" is still misaligned.
For example, in Warehouse_0000, Camera_0001 and Camera_0026 are misaligned. (at 00:04:57)
Hi
@visionNoob
. We have updated Warehouse_000 and removed the misaligned video sequences. Please let us know if you observe any other issues. Thanks.