Modes Overview

The engine exposes one command group. Domain components remain usable on their own and pipelines contain no source, output, persistence, retry, or display logic.

Mode	Purpose	Required input
install	Install missing optional dependencies into the current Python environment	--extra or --requirement
track	Track a source or replay saved TrackR-CNN predictions	--source plus component selectors, or --detections, --images, and --instances
materialize	Publish keyed perception or derive a timestamped frame-loss dataset	experiment, or --time-variant with dataset, sequence, and --build
eval	Evaluate a perception build or saved sensor inputs	experiment, dataset plus --build, or sensor dataset plus --tracker eagermot
tune	Optimize tracker parameters from a build or saved sensor inputs	experiment, dataset plus --build, or sensor dataset plus --tracker eagermot
research	Score proposed tracker changes against a build	experiment plus --build
train-reid	Train a reusable appearance backbone	ReID dataset/config
eval-reid	Evaluate query/gallery retrieval	checkpoint and ReID dataset
compare-reid	Compare checkpoints across ReID datasets	checkpoints and targets
export	Export a ReID backbone	checkpoint and formats
build	Compile native tracker libraries	tracker selector/toolchain

Live tracking

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

The engine creates canonical RGB Frame values, runs a TrackingPipeline, and sends each (frame, result) pair to configured sinks.

Reproducible benchmark workflow

Perception runs once during materialization. Eval and tune can perform that deterministic step automatically; research consumes an explicit build:

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

These workflows never select a latest build. They verify source, split, taxonomy, geometry, and component fingerprints before replay.

See Materialize, Evaluate, Tune, and Research.

Use materialize --time-variant to compare timing modes under reproducible frame loss without rerunning perception.

Python composition

Python callers use factories and canonical structures directly. See the Python API for detector, segmentor, appearance encoder, tracker, and pipeline examples.