One API, nine trackers and a reason to stop rewriting MOT boilerplate

What it does

BoxMOT is a unified interface for modern multi-object tracking. It exposes one CLI and one Python API that covers tracking, benchmark evaluation, hyperparameter tuning, research loops, and ReID model training. The project wraps multiple recent trackers—including BoT-SORT, ByteTrack, OC-SORT, and StrongSORT—behind shared detector and ReID plumbing, so experiments share the same preprocessing and metrics instead of fragmented scripts.

The interesting bit

The optional C++ tracker backend is the unusual angle: it promises the same benchmark metrics as the Python path and can be embedded in standalone C++ projects via CMake. It also handles both axis-aligned and oriented bounding boxes, which matters when objects rotate and upright rectangles waste pixels.

Key highlights

• Swappable trackers with cached detections and embeddings for benchmark workflows.
• Nine trackers compared on MOT17, SportsMOT, and MMOT benchmarks (see the README table).
• Native C++ implementations available for some trackers; metric parity is claimed.
• Supports both AABB and OBB tracking paths.
• AGPL 3.0 license; commercial support is advertised.

Caveats

• C++ backend coverage is spotty: the benchmark table shows many trackers lack C++ scores entirely, and several lack MMOT or SportsMOT results, so cross-dataset comparisons are sparse.

Verdict

Use this if you are a researcher or engineer who needs to compare or productionize tracking algorithms without maintaining nine separate codebases. Skip it if you only need one specific tracker and prefer to minimize abstraction layers.