How mikel-brostrom/boxmot works
Project type and purpose
• BoxMOT is an open-source Python library and CLI tool for multi-object tracking (MOT).
• It provides a pluggable, state-of-the-art tracking framework that integrates seamlessly with detection, segmentation, and pose models and a zoo of ReID backbones.
• Users interact via a unified CLI (boxmot) that supports subcommands for tracking, generating detections and embeddings, evaluation, tuning, and exporting ReID models.
File structure
• Top-level entry points:
• Dockerfile, pyproject.toml and manifest for packaging.
• README.md and LICENSE.
• boxmot/engine/cli.py — CLI interface.
• Modular library layout under boxmot/:
• appearance/ — ReID backbones, exporters, auto-backends.
• configs/ — YAML configs for each tracker method.
• data/ — dataset loader and utilities.
• engine/ — orchestrates detection, tracking, evaluation, tuning.
• motion/ — motion models and Kalman filters.
• postprocessing/ — output smoothing (GSI).
• trackers/ — implementations of SOTA trackers (boosttrack, botsort, etc.).
• utils/ — common helpers (association, IOU, plotting).
• Example notebooks under examples/.
• Tests under tests/ for unit and performance.
README for additional insights
• Architecture is pipeline-driven:
Input source → Detector → (optional) ReID embedding → Tracker → Postprocessing → Output/evaluation.
• Supports multiple backends (PyTorch, ONNX, TensorRT, OpenVINO) for ReID export and inference.
• Configs dictate hyperparameters for detectors, trackers, and motion models.