feature: ReID support for bot sort tracker

[zhiweit]

Hi, thank you for the community support. I would like to ask if there is plan for adding ReID model with cuda support for bot sort tracker in the future? It would be great if it has this support

[JESUSROYETH]

Hi @zhiweit 👋

I can't speak for the maintainers about the roadmap, but I can share the current state of the library, it may help you decide what to do today.

The current BotSortTracker is motion-only: Kalman filter plus camera motion compensation, there is no appearance/ReID branch. The ReID infrastructure did exist in the package: DeepSORTTracker shipped with a feature extractor that runs on GPU, and it was removed in 2.1.0 as a breaking change (see the CHANGELOG: "Projects using DeepSort must pin to <2.1.0").

If you need appearance features with CUDA today, pinning the old version still works. One warning: the supervision pin matters. trackers 2.0.1 imports box_iou_batch from a path that changed in supervision 0.26.0 stable, so the only combination I got working is:

pip install "trackers[deepsort]==2.0.1" "supervision==0.26.0rc7"

from trackers import DeepSORTFeatureExtractor, DeepSORTTracker

# timm >= 1.0 needs the full model name with tag, plain "resnet18" is rejected
extractor = DeepSORTFeatureExtractor.from_timm("resnet18.a1_in1k", device="cuda")
tracker = DeepSORTTracker(feature_extractor=extractor)

tracked = tracker.update(detections, frame)  # the frame is used for the ReID crops

I verified this locally: the extractor loads on cuda (~55 MB of GPU memory with resnet18) and the IDs stay stable across frames. Versions: trackers 2.0.1, supervision 0.26.0rc7, timm 1.0.27, torch 2.12.1+cu130, Python 3.12, RTX 4060.

Hope this helps!

[AlexBodner]

Hey @zhiweit!
BoTSORT with ReID is coming soon to the library, stay tuned!