supervision

Why we included this project

Most computer vision projects end up with the same hand-rolled code: parsing model outputs, drawing boxes, tracking objects between frames, counting detections in zones. Supervision wraps that work into a model-agnostic library, so you can point it at predictions from Ultralytics, Detectron2, MMDetection, SAM, or a vision-language parser and get consistent annotation, filtering, and metrics without rewriting anything. The unified Detections API is what makes the swap painless: change the underlying model and your downstream code stays put. If you are moving a prototype toward real-time zone counting or video processing, the building blocks are already here. It is a solid layer to standardize on when your stack mixes several detection frameworks.