boxmot

What it is and what it does

BoxMOT is a multi-object tracking framework that bridges detection and tracking workflows. It provides a unified interface for running tracking experiments, benchmarking, ReID training, and production inference across multiple tracker implementations (OccluBoost, ByteTrack, BoTSORT, and others). The package handles both axis-aligned and oriented bounding box tracking, supports swappable detector and ReID components, and offers both a Python API for embedding in applications and a CLI for standalone use.

The core workflow is tracking-by-detection: you feed it detections (bounding boxes with confidence scores) frame-by-frame, and it assigns consistent IDs across frames using motion and appearance cues. It includes benchmark-oriented features like cached detection/embedding reuse to avoid redundant computation across experiments, native C++ tracker implementations for production deployment, and integrated ReID training pipelines. Dependencies are substantial (torch, torchvision, opencv-python, scikit-learn, timm) but standard for computer vision; the package itself is pure Python and installs with low friction.

Use it for:

• Track pedestrians, vehicles, or sports players across video frames using detections from YOLO or another detector.
• Benchmark tracker performance on standard datasets (MOT17, SportsMOT, MMOT) with cached workflows to avoid recomputing detections.
• Train and evaluate ReID (re-identification) models to improve tracking accuracy on specific object classes or domains.
• Embed multi-object tracking in a Python application or Jupyter notebook without building a custom tracker from scratch.
• Export a production-ready C++ tracker implementation for deployment in standalone applications via CMake.
• Experiment with different tracker algorithms and hyperparameters using the CLI or Python API without modifying detector/ReID code.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

BoxMOT provides pluggable multi-object tracking modules that work with bounding box detections from any model, supporting both axis-aligned and oriented bounding boxes through a unified Python API and CLI.

Yes, with AGPL-3.0 license caveat. BoxMOT is actively maintained, has low install friction, and offers a comprehensive, well-benchmarked tracking framework suitable for research, benchmarking, and production use. The AGPL-3.0 license requires that any modifications or derivative works be open-source; if you need proprietary tracking or closed-source modifications, this is not the right choice. For academic research, open-source projects, or applications where AGPL-3.0 compliance is acceptable, it is worth installing.

Install

boxmot on PyPI

pip install boxmot

uv add boxmot

poetry add boxmot

Installing boxmot

Before you install

Low install friction with a pure-Python wheel. Actively maintained with a recent release 35 days ago. Supports Python 3.10 through 3.13. Depends on 13 runtime packages including torch, torchvision, and opencv-python, which are substantial but standard for computer vision work.

License in practice

Licensed under AGPL-3.0, which requires that any modifications or derivative works distributed must also be open-source under AGPL-3.0. This is a strong copyleft license; proprietary use or closed-source modifications are not permitted without explicit permission.

Quickstart

pip install boxmot

import numpy as np
from boxmot.trackers import OccluBoost

tracker = OccluBoost()
dets = np.array([[100, 200, 300, 400, 0.9, 0]], dtype=np.float32)
img = np.zeros((480, 640, 3), dtype=np.uint8)
tracks = tracker.update(dets, img)

Requires torch and torchvision; GPU recommended for real-time performance. Pre-trained ReID and tracker models download on first use via gdown and huggingface-hub.

Verify before relying

• Whether native C++ tracker backend (--tracker-backend cpp) requires separate compilation or system dependencies beyond what pip provides.
• Whether mode-specific extras (yolo, evolve, research, onnx, openvino, tflite) are documented and what additional dependencies they pull in.
• Typical inference latency and throughput for common tracker/detector combinations on standard hardware.

Package facts

Evidence: boxmot-22.0.0-py3-none-any.whl

Keywords: AI, DL, ML, YOLO, deep-learning, machine-learning, tracking, tracking-by-detection, vision

Tags

More Software Development packages