SparseQID

Online multi-camera 3D tracking by ID prediction over recurrent sparse queries.

SparseQID combines a recurrent outside-in Sparse4D detector with an online identity-prediction module. The detector fuses synchronized, calibrated cameras into world-frame 3D observations. SparseQIDTracker assigns persistent scene-global IDs using the detector's query features and a finite trajectory memory.

The system placed third in Track 1 of the 2026 AI City Challenge. Its identity layer raised test HOTA from 29.63 to 38.01 over native detector identities, primarily through an AssA increase from 20.83 to 31.10.

Install and test

Python 3.11 or 3.12 and a CUDA-capable GPU are recommended.

uv sync --dev
uv run pytest

The package exposes one command with four subcommands:

sqid extract     build the JPEG frame cache
sqid train       train the identity model with a frozen recurrent detector
sqid infer       run detector, identity assignment, and submission writing
sqid visualize   render tracked boxes as MP4 clips

• docs/QUICKSTART.md — nothing to a tracking video on one validation scene, in six commands.
• docs/REPRODUCE.md — dataset download, expected directory layout, frame caches, training, inference, and evaluation, end to end.
• docs/CLI.md — every option of the four subcommands, with defaults.

The data comes from NVIDIA's PhysicalAI-SmartSpaces dataset (subsets MTMC_Tracking_2026 and MTMC_Tracking_2025); it is public and needs no Hugging Face login.

Checkpoints

Paper weights are hosted at playbox-dev/SparseQID on Hugging Face. Their experimental roles and test results are listed in docs/CHECKPOINTS.md. Download them with:

uvx --from huggingface-hub hf download playbox-dev/SparseQID \
  --include "checkpoints/**" --local-dir .

Each checkpoint contains the backbone, neck, detector head, trajectory memory, identity decoder, and position encoder needed for inference.

Evaluation

SparseQID does not redistribute or wrap an evaluator. NVIDIA’s current Physical AI Smart Spaces dataset documentation points to the official offline 3D-box HOTA implementation, evaluate_aicity_mtmc.py.

Citation

@inproceedings{shrestha2026sparseqid,
  title     = {Online Multi-Camera 3D Tracking via ID Prediction over Recurrent Sparse Queries},
  author    = {Shrestha, Pragyan and Nakayama, Haruto and Scott, Atom},
  booktitle = {ECCV Workshops},
  year      = {2026}
}

Please also cite the MOTIP and outside-in Sparse4D papers when using their corresponding components.

License

SparseQID includes code adapted from NVIDIA TAO Sparse4D and MOTIP. Original copyright headers are retained in adapted files.

See LICENSE and NOTICE.