Supervision

Supervision is Roboflow’s open-source Python library with 51,200+ stars on GitHub — a model-agnostic toolkit of reusable computer vision utilities for detection, segmentation and tracking.

Overview

Every vision project rewrites the same glue code: drawing boxes, converting dataset formats, computing metrics, counting objects as they cross a line. Supervision packages that glue once, in a well-maintained library. It accepts output from virtually any model — YOLO, DETR, Segment Anything, Hugging Face transformers and more — through a unified Detections data format, then covers the rest of the pipeline: loading and converting datasets in COCO, YOLO and Pascal VOC formats, a rich set of annotators, ByteTrack multi-object tracking, polygon zones and line counters for analytics, inference slicing for small objects, and evaluation metrics such as mAP and confusion matrices. Maintained by Roboflow and a large community, it installs with pip and works alongside PyTorch, TensorFlow or ONNX Runtime rather than replacing them. The repository is updated continuously, most recently in October 2026.

Key Features

• Model-agnostic Detections API – one format for output from YOLO, SAM, transformers and other models.
• Rich annotators – boxes, masks, labels, traces, heat maps and more for instant visualization.
• Dataset tools – load, split, merge and convert between COCO, YOLO and Pascal VOC formats.
• ByteTrack tracking – production-ready multi-object tracking built in.
• Zones and line counters – count and analyze objects in regions or crossing lines for video analytics.
• Evaluation metrics – mean average precision, F1, recall and confusion matrices out of the box.
• Inference slicing – detect small objects in large images by slicing them into tiles.

💳 Pricing

Supervision is completely free and open source under the MIT license. It is a library you install into your own projects; Roboflow’s separate hosted platform is optional and not required to use any part of Supervision.

Who Is It For

Computer vision engineers, data scientists and video-analytics builders who want production-grade utilities for detection and tracking pipelines without reinventing the same plumbing on every project.