BoxMOT

BoxMOT is a structured tracking-by-detection toolkit for axis-aligned boxes, oriented boxes, and instance masks. Its dependency direction is explicit:

structures -> domain components -> pipelines -> engine

Detectors, segmentors, appearance encoders, and trackers are independently usable. Pipelines normally provide reusable masks and appearance embeddings. For live use, every high-level ReID-enabled tracker adapter can also extract missing embeddings lazily from a supplied frame. Native adapters pass those features to their model-free C++ tracker libraries. The engine owns media sources, sinks, services, CLI workflows, evaluation, tuning, and resumable materialization.

MafHda combines motion and masked correlation-filter appearance for AABB detections with nonempty full-frame instance masks and current image frames. boxmot track --tracker maf_hda --detections DIR --images DIR --instances DIR replays saved TrackR-CNN predictions on KITTI MOTS; see the MAF-HDA tracker page for the required paths. EagerMot adds 2D/3D sensor fusion through the Python tracker API with independent detection batches and camera calibration. Use boxmot eval --tracker eagermot to evaluate saved KITTI PointGNN and TrackR-CNN predictions against KITTI MOTS masks; the tracker page provides the command. boxmot tune --dataset ./kitti-mots --tracker eagermot loads a multimodal sequence dataset.yaml, optimizes separate car and pedestrian profiles together for class-average mask HOTA, and saves profiles for boxmot eval --tracker eagermot --dataset ./kitti-mots --class-config. Sequence data lives together; the same YAML declares classes, splits, and the encoding and paths for each modality. Use the sensor dataset template to supply your own images, 2D/3D detections, calibration, and ego poses.

Get started

pip install boxmot
boxmot --help

boxmot track \
  --source video.mp4 \
  --detector yolov8n \
  --tracker bytetrack \
  --save

For repeatable evaluation, publish perception once and name the build on every downstream command:

boxmot materialize --experiment mot17/ablation-yolox-lmbn.yaml
boxmot eval --experiment mot17/ablation-yolox-lmbn.yaml --build BUILD_ID

Where to go next

Goal	Guide
Install optional detector, service, or export runtimes	Installation
Run the CLI	CLI
Track live or finite media	Track
Build reusable keyed perception data	Materialize
Evaluate, tune, or research against a fixed build	Modes
Compose Torch-native components in Python	Python API
Select an association algorithm	Trackers
Embed a service or native tracker	Integrations

Canonical structures are CPU-contiguous and validated without implicit conversion. Standalone box-only tracker calls may use exact NumPy AABB6 or OBB7 rows and receive packed NumPy AABB8 or OBB9 rows; pipelines and enriched detections use canonical structures.