BoxMOT: Pluggable SOTA multi-object tracking modules modules for segmentation, object detection and pose estimation models

🚀 Key Features

• Pluggable Architecture
  Easily swap in/out SOTA multi-object trackers.

• Universal Model Support
  Integrate with any segmentation, object-detection and pose-estimation backbones that outputs bounding boxes

• Benchmark-Ready
  Local ealuation pipelines for MOT17, MOT20, and DanceTrack ablation datasets with "official" ablation detectors

• Performance Modes

  • Motion-only: for lightweight, CPU-efficient, high-FPS performance
  • Motion + Appearance: Combines motion cues with appearance embeddings (CLIPReID, LightMBN, OSNet) to maximize identity consistency and accuracy at a higher computational cost

• Reusable Detections & Embeddings
  Save once, run any tracker with no redundant preprocessing.

📊 Benchmark Results (MOT17 ablation split)

Tracker	Status	HOTA↑	MOTA↑	IDF1↑	FPS
boosttrack	✅	69.253	75.914	83.206	25
botsort	✅	68.885	78.222	81.344	46
strongsort	✅	68.05	76.185	80.763	17
deepocsort	✅	67.796	75.868	80.514	12
bytetrack	✅	67.68	78.039	79.157	1265
ocsort	✅	66.441	74.548	77.899	1483

_{NOTES: Evaluation was conducted on the second half of the MOT17 training set, as the validation set is not publicly available and the ablation detector was trained on the first half. We employed pre-generated detections and embeddings. Each tracker was configured using the default parameters from their official repositories.}

🔧 Installation

Install the boxmot package, including all requirements, in a Python>=3.9 environment:

pip install boxmot

BoxMOT provides a unified CLI boxmot with the following subcommands:

Usage: boxmot COMMAND [ARGS]...

Commands:
  track                  Run tracking only
  generate               Generate detections and embeddings
  eval                   Evaluate tracking performance using the official trackeval repository
  tune                   Tune tracker hyperparameters based on selected detections and embeddings

If you want to contribute to this package check how to contribute here

📝 Code Examples & Tutorials

Tracking

$ boxmot track --yolo-model rf-detr-base.pt     # bboxes only
  boxmot track --yolo-model yolox_s.pt          # bboxes only
  boxmot track --yolo-model yolo12n.pt         # bboxes only
  boxmot track --yolo-model yolo11n.pt         # bboxes only
  boxmot track --yolo-model yolov10n.pt         # bboxes only
  boxmot track --yolo-model yolov9c.pt          # bboxes only
  boxmot track --yolo-model yolov8n.pt          # bboxes only
                            yolov8n-seg.pt      # bboxes + segmentation masks
                            yolov8n-pose.pt     # bboxes + pose estimation

Tracking methods

$ boxmot track --tracking-method deepocsort
                                 strongsort
                                 ocsort
                                 bytetrack
                                 botsort
                                 boosttrack

Tracking sources

Tracking can be run on most video formats

$ boxmot track --source 0                               # webcam
                        img.jpg                         # image
                        vid.mp4                         # video
                        path/                           # directory
                        path/*.jpg                      # glob
                        'https://youtu.be/Zgi9g1ksQHc'  # YouTube
                        'rtsp://example.com/media.mp4'  # RTSP, RTMP, HTTP stream

Select ReID model

Some tracking methods combine appearance description and motion in the process of tracking. For those which use appearance, you can choose a ReID model based on your needs from this ReID model zoo. These model can be further optimized for you needs by the reid_export.py script

$ boxmot track --source 0 --reid-model lmbn_n_cuhk03_d.pt               # lightweight
                                       osnet_x0_25_market1501.pt
                                       mobilenetv2_x1_4_msmt17.engine
                                       resnet50_msmt17.onnx
                                       osnet_x1_0_msmt17.pt
                                       clip_market1501.pt               # heavy
                                       clip_vehicleid.pt
                                      ...

Filter tracked classes

By default the tracker tracks all MS COCO classes.

If you want to track a subset of the classes that you model predicts, add their corresponding index after the classes flag,

boxmot track --source 0 --yolo-model yolov8s.pt --classes 16 17  # COCO yolov8 model. Track cats and dogs, only

Here is a list of all the possible objects that a Yolov8 model trained on MS COCO can detect. Notice that the indexing for the classes in this repo starts at zero

Evaluation

Evaluate a combination of detector, tracking method and ReID model on standard MOT dataset or you custom one by

# reproduce MOT17 README results
$ boxmot eval --yolo-model yolox_x_MOT17_ablation.pt --reid-model lmbn_n_duke.pt --tracking-method boosttrack --source MOT17-ablation --verbose 
# MOT20 results
$ boxmot eval --yolo-model yolox_x_MOT20_ablation.pt --reid-model lmbn_n_duke.pt --tracking-method boosttrack --source MOT20-ablation --verbose 
# Dancetrack results
$ boxmot eval --yolo-model yolox_x_dancetrack_ablation.pt --reid-model lmbn_n_duke.pt --tracking-method boosttrack --source dancetrack-ablation --verbose 
# metrics on custom dataset
$ boxmot eval --yolo-model yolov8n.pt --reid-model osnet_x0_25_msmt17.pt --tracking-method deepocsort  --source ./assets/MOT17-mini/train --verbose

add --gsi to your command for postprocessing the MOT results by gaussian smoothed interpolation. Detections and embeddings are stored for the selected YOLO and ReID model respectively. They can then be loaded into any tracking algorithm. Avoiding the overhead of repeatedly generating this data.

Evolution

We use a fast and elitist multiobjective genetic algorithm for tracker hyperparameter tuning. By default the objectives are: HOTA, MOTA, IDF1. Run it by

# saves dets and embs under ./runs/dets_n_embs separately for each selected yolo and reid model
$ boxmot generate --source ./assets/MOT17-mini/train --yolo-model yolov8n.pt yolov8s.pt --reid-model weights/osnet_x0_25_msmt17.pt
# evolve parameters for specified tracking method using the selected detections and embeddings generated in the previous step
$ boxmot tune --dets yolov8n --embs osnet_x0_25_msmt17 --n-trials 9 --tracking-method botsort --source ./assets/MOT17-mini/train

The set of hyperparameters leading to the best HOTA result are written to the tracker's config file.

Export

We support ReID model export to ONNX, OpenVINO, TorchScript and TensorRT

# export to ONNX
$ python3 boxmot/appearance/reid_export.py --include onnx --device cpu
# export to OpenVINO
$ python3 boxmot/appearance/reid_export.py --include openvino --device cpu
# export to TensorRT with dynamic input
$ python3 boxmot/appearance/reid_export.py --include engine --device 0 --dynamic

Example Description	Notebook
Torchvision bounding box tracking with BoxMOT
Torchvision pose tracking with BoxMOT
Torchvision segmentation tracking with BoxMOT

Contributors

Contact

For BoxMOT bugs and feature requests please visit GitHub Issues. For business inquiries or professional support requests please send an email to: box-mot@outlook.com