SparseQID checkpoints

Checkpoints for Online Multi-Camera 3D Tracking via ID Prediction over Recurrent Sparse Queries. Source code and instructions are available at playbox-dev/SparseQID.

The submitted model is ngc-frozen_v128_pe30_it6k.pth. It achieved 38.01 HOTA on the 2026 AI City Challenge Track 1 test set.

Download

Run from the SparseQID source repository:

hf download playbox-dev/SparseQID \
  --include "checkpoints/**" \
  --local-dir .

Files

Checkpoint	Use
ngc-frozen_v128_pe30_it6k.pth	Submitted model; frozen NGC detector
ngc-joint_v128_pe30_it5k.pth	Joint detector and identity training
ngc-ft4k-frozen_v128_pe30_it8k.pth	Frozen finetuned detector
ngc-ft4k-joint_v128_pe30_it2k.pth	Joint training with finetuned detector
tok-appearance_v128_pe30_it2k.pth	Appearance-token ablation
tok-pos-fourier_v128_pe30_it2k.pth	Fourier-position-token ablation
tok-pos-raw_v128_pe30_it2k.pth	Raw-position-token ablation

License

See LICENSE and NOTICE.