Most multi-object tracking in production is a graft: one research fork for ByteTrack, another for BoT-SORT, and a layer of glue translating each one's bespoke detection format. This library takes the opposite bet — every algorithm is re-implemented clean-room from its paper against a single API, so switching from SORT to BoT-SORT is a one-line change instead of a rewrite.
What Sets It Apart
- Detector-agnostic by design: it consumes
supervision.Detections, so YOLO, DETR, RT-DETR, or any box producer plugs in with no adapter code — the tracker never cares where the boxes came from. - One interface spanning SORT, ByteTrack, OC-SORT, BoT-SORT, and C-BIoU, which means you can A/B algorithms on the same pipeline rather than maintaining five separate integrations.
- Tuning and benchmarking are built in, not left as homework: Optuna-driven hyperparameter search (
trackers tune) plus an evaluation suite reporting CLEAR, HOTA, and Identity metrics across MOT17, SportsMOT, SoccerNet, and DanceTrack. - BoT-SORT adds camera-motion compensation, so moving-camera footage from drones or sports broadcasts doesn't break association the way fixed-camera trackers do.
Who It's For
Great fit if you need to compare trackers on your own detector, or want a maintained, readable reference implementation instead of a half-abandoned paper repo. The reported HOTA spread of 47.2–84.5 across datasets makes it a sensible bench for choosing an algorithm before you commit. Look elsewhere if you need a single battle-hardened tracker squeezed for one domain — a purpose-built fork may still edge it on raw numbers — or if your problem needs trackers outside the motion-and-appearance family it covers.
