nct-tso/STIR_pseudo_tracks

STIROrig multi-teacher point tracks

This dataset contains cached trajectories produced by six point trackers on STIROrig. It contains no source video frames. The archive is the phase-1 output used by the verifier-guided pseudo-label pipeline in https://github.com/danushkv/STIR2026_challenge.

Dataset structure

<teacher>/<patient>/<side>__<seq>__<teacher>.npz

The teachers are alltracker, cotracker3, locotrack, mft, trackon2, and trackon_r. Coverage can differ because some teachers skipped clips they could not process; downstream code groups clips by the set of available teachers.

teacher	files
alltracker	487
cotracker3	566
locotrack	487
mft	487
trackon2	487
trackon_r	487
total	3,001

The current directory occupies approximately 36 MiB before Hugging Face storage overhead.

Each compressed NumPy file contains:

key	shape	meaning
fwd_coords	[N, T, 2]	forward (x, y) tracks in native-image pixels
fwd_vis	[N, T]	forward visibility score
fwd_conf	[N, T]	forward teacher confidence
name	scalar string	teacher identifier
bwd_coords	[N, T, 2]	backward/cycle track, when available
bwd_vis	[N, T]	backward visibility, when available
bwd_conf	[N, T]	backward confidence, when available

N is the number of segmentation-derived query points and T is the number of retained frames. Tracks were collected with --skip 5; clip identifiers and frame selection must remain unchanged when regenerating pseudo-labels.

Download and use

hf download nct-tso/STIR_pseudo_tracks \
  --repo-type dataset --local-dir data/STIROrig_tracks

The files can be loaded with numpy.load. The repository code provides load_track_result_pair in src/trajectories.py and the complete fusion entry point in src/run_phase2.py. manifest.csv records the size and SHA-256 of every file and can be regenerated with tools/build_tracks_manifest.sh.

How the tracks were produced

The code repository records the exact command and teacher adapters. In short:

for teacher in cotracker3 alltracker locotrack trackon2 trackon_r mft; do
  bash scripts/collect_teacher.sh "$teacher" orig
done

Queries come from the first-frame IR tattoo segmentation. Every teacher runs both a forward and, where supported, backward pass. Model-specific confidence and visibility values are retained without calibration. No manual corrections were applied to these raw files.

Intended use

These files are intended to reproduce or study multi-teacher verification, pseudo-label generation, and point-tracking distillation without rerunning all six expensive teacher models. They are not ground truth. Do not use their predictions as clinical annotations or for patient-level decisions.

Biases and limitations

• Teacher errors and biases are preserved in the raw trajectories.
• STIR provides endpoint annotations, not dense per-frame ground truth.
• Visibility and confidence scales differ between teachers.
• Users still need lawful access to STIROrig to decode videos, inspect tracks, or run the verifier.

Licence and source-data terms

The metadata deliberately uses license: other while redistribution of derived coordinates is reviewed against the STIROrig terms and the relevant teacher model licences. Access to the original videos is governed by STIROrig's own terms; those videos are not included here.

Citation

Please cite both this release and the original STIROrig dataset:

@software{venkatesh2026stir,
  author  = {Venkatesh, Danush Kumar and Liu, Peng and Speidel, Stefanie},
  title   = {Verifier-Guided Multi-Teacher Distillation for Streaming Tissue Tracking},
  year    = {2026},
  url     = {https://github.com/danushkv/STIR2026_challenge},
  note    = {STIR Challenge 2026, Team NCT\_TSO}
}

@article{schmidt2024stir,
  author  = {Schmidt, Adam and Mohareri, Omid and DiMaio, Simon P. and Salcudean, Septimiu E.},
  title   = {Surgical Tattoos in Infrared: A Dataset for Quantifying Tissue
             Tracking and Mapping},
  journal = {IEEE Transactions on Medical Imaging},
  volume  = {43},
  number  = {7},
  pages   = {2634--2645},
  year    = {2024},
  doi     = {10.1109/TMI.2024.3372828}
}