roboflow/supervision

Supervision is a Python library from Roboflow that provides reusable building blocks for computer vision applications. Computer vision means teaching computers to interpret images and video, identifying objects, tracking movement, measuring areas of interest, and so on. Writing the same boilerplate code for drawing bounding boxes, counting detections, managing zones, and handling video streams is tedious, and Supervision solves that by providing those tools as ready-made, well-tested components. The library is model-agnostic, meaning it works with output from any detection or segmentation model, whether that is Ultralytics YOLO, Hugging Face Transformers, MMDetection, or Roboflow's own Inference service. You get the model's raw output, convert it into Supervision's standard Detections format with a single function call, and then use Supervision's tools to annotate images, track objects across video frames, count how many detections pass through a defined zone, measure speed, or save the results. It also includes utilities for loading, splitting, merging, and converting datasets between popular formats like COCO, YOLO, and Pascal VOC. You would use Supervision when building any project that takes a computer vision model's output and needs to do something useful with it, for example, a security camera system that counts people entering a store, a traffic monitoring tool that estimates vehicle speed, or a quality-control system that flags defective products on a conveyor belt. Rather than writing custom annotation and tracking logic from scratch, you import Supervision and connect it to whichever model you are using. It runs on Python 3.9 and later and installs via pip.