Trackers is a Python library that provides clean, modular implementations of leading multi-object tracking algorithms designed to work with any detection model.
The library solves the problem of integrating object tracking into detection pipelines by offering re-implemented versions of algorithms like SORT, ByteTrack, OC-SORT, BoT-SORT, C-BIoU, and McByte. Rather than wrapping or vendoring existing code, each algorithm is reimplemented from its original paper, making the implementations readable and modifiable. The tool works natively with the supervision.Detections format, allowing it to slot directly into pipelines using YOLO, DETR, RT-DETR, or any detector that produces bounding boxes without requiring additional glue code.
Adoption makes sense for teams building production tracking systems, researchers comparing algorithms, or developers who need detector-agnostic tracking without licensing constraints. The Apache 2.0 license permits use in closed-source products, distinguishing it from copyleft alternatives. The library includes benchmarked implementations across four datasets—MOT17, SportsMOT, SoccerNet, and DanceTrack—both at default parameters and after hyperparameter tuning, so performance expectations are clear before deployment. For scene-specific optimization, an optional hyperparameter tuning feature uses Optuna to search for ideal settings. BoT-SORT and McByte handle moving cameras natively, maintaining stable track IDs when the frame shifts.
Development activity shows consistent attention to practical usability: the project maintains a command-line interface for tracking video, webcam, RTSP streams, or image directories without writing Python code, and documentation covers both programmatic and CLI workflows. The codebase emphasizes transparency by keeping implementations readable and modifiable rather than obscured, and the benchmarking work across multiple datasets reflects a commitment to setting accurate performance expectations.
