Choose a Mode

BoxMOT has two main usage styles:

• direct source tracking for videos, cameras, image folders, and streams
• config-driven benchmark modes for repeatable experiments

Mode Guide

Mode	Use it when	Main command	Install notes	Guide
Track	You want detector + tracker output on a live or saved source	boxmot track	Core install is enough if detector backends are already available. Install yolo upfront for common YOLO workflows.	Track
Generate	You want reusable detections and embeddings	boxmot generate	Same detector-backend requirement as track. yolo is the common extra to preinstall.	Generate
Evaluate	You want TrackEval metrics on a benchmark	boxmot eval	Usually the same detector-backend setup as generate, because benchmark evaluation reuses generated detections and embeddings.	Evaluate
Tune	You want to search tracker hyperparameters	boxmot tune	Add the evolve extra.	Tune
Research	You want GEPA to propose tracker code changes	boxmot research	Add the research extra.	Research
Export	You want to convert a ReID model to deployment formats	boxmot export	Add format-specific extras such as onnx, openvino, or tflite.	Export

See Installation for exact commands.

Two common paths

Direct source path

Use this when you already have a source and want outputs quickly:

boxmot track --detector yolov8n --reid osnet_x0_25_msmt17 --tracker botsort --source video.mp4 --save

Benchmark path

Use this when you want repeatable experiments on MOT-style datasets:

boxmot generate --benchmark mot17-ablation
boxmot eval --benchmark mot17-ablation --tracker boosttrack
boxmot tune --benchmark mot17-ablation --tracker bytetrack

In this path, the benchmark config resolves the dataset, detector profile, and ReID profile automatically.

Python API path

If you want the same modes in Python, start with Python API Overview. The public facade is boxmot.Boxmot.