Skip to content

Releases: Athlitix/arcsort

Release list

ARC-SORT v0.1.1

ARC-SORT v0.1.1 Latest
Latest

Choose a tag to compare

View all tags
@agirishkumar agirishkumar released this 22 Oct 21:28
v0.1.1
3f7ca1c
  • Output boxes: tracks now report the detection box they were matched to this frame
    (output_box="detection", new default). output_box="kalman" keeps the 0.1.0 behaviour (the
    Kalman-filtered state). Association is unchanged; HOTA rises in every setting tested, e.g.
    validation with the full pipeline: SportsMOT 84.957 → 86.295, DanceTrack 60.617 → 61.329.
    scripts/reproduce.sh keeps OUTPUT_BOX=kalman by default so arXiv:2610.22306 v1 still
    reproduces exactly; scripts/regression_gate.sh stays byte-identical on all three submissions.
  • Motion-compensated trails: Visualizer(motion_compensated_trails=True) with the new
    ARCTracker.camera_motion property; tools/pipeline.py --motion-trails. Uses Roboflow
    trackers (now part of the optional viz extra).
  • tools/run_cached.py --output-box; tools/evaluate.py --benchmark soccernet.
  • README: comparison with Roboflow trackers 2.6.1 (SoccerNet test, SportsMOT/DanceTrack val with
    detections and GT boxes); tested-detectors table rerun with 0.1.1.

pip install -U arcsort · Project page · Paper

Assets 2
Loading

ARC-SORT v0.1.0

ARC-SORT v0.1.0

Choose a tag to compare

View all tags
@agirishkumar agirishkumar released this 17 Oct 15:00
v0.1.0
2a580c5

First public release of ARC-SORT, the code for arXiv:2610.22306.

pip install arcsort

Reproduction files

  • arcsort-v0.1.0-caches.tar.gz: our cached YOLOX-X detections and camera-motion transforms for the SportsMOT and DanceTrack val/test splits. Extract at the repo root, then run scripts/reproduce.sh (see the README).
  • sportsmot_test.zip (81.82 HOTA), dancetrack_test.zip (62.1 HOTA), sportsmot_deepeiou_test.zip (80.31 HOTA, on Deep-EIoU's inputs): the result files we submitted to the CodaBench evaluation servers. scripts/regression_gate.sh checks a run against them.

Detector and ReID weights are not redistributed; the README links to the original authors' releases.

Highlights

  • Training-free tracker: IMM Kalman filter, uncertainty-adaptive Expansion IoU, camera-motion compensation, BYTE, deferred identity arbitration, and an appearance-only offline merge.
  • ARCTracker.from_preset("sportsmot" | "dancetrack"); the core is CPU-only and needs no torch.
  • Tested with YOLOv8/v10/11/12/26/v13, RT-DETRv2, RF-DETR and YOLOX detectors.
Loading