Supervision

Supervision is an open-source Python library by Roboflow that provides reusable, model-agnostic tools for building computer vision applications. It supports loading datasets, drawing detections on images and videos, counting objects in zones, and much more. Designed to work with any classification, detection, or segmentation model, Supervision integrates seamlessly with popular frameworks like Ultralytics, Hugging Face Transformers, MMDetection, and Roboflow Inference.

• Model-Agnostic Connectors: Plug in any detection, classification, or segmentation model via built-in connectors for Ultralytics, Transformers, MMDetection, Inference, and rfdetr — just call sv.Detections.from_* to get started.
• Highly Customizable Annotators: Compose rich visualizations using a wide range of annotators (bounding boxes, masks, labels, etc.) by instantiating an annotator class and calling .annotate() on your image or frame.
• Dataset Utilities: Load, split, merge, and save datasets in YOLO, COCO, and Pascal VOC formats using sv.DetectionDataset — enabling easy format conversion and dataset management.
• Object Tracking: Integrate multi-object tracking (e.g., ByteTrack) to follow detections across video frames with minimal setup.
• Zone-Based Analytics: Define polygonal zones and count or analyze detections within them, enabling use cases like dwell time analysis and crowd monitoring.
• Video Processing: Process video streams frame-by-frame using sv.VideoInfo and sv.VideoSink utilities for real-time or batch inference pipelines.
• Speed & Metrics: Estimate object speed, compute detection metrics, and evaluate model performance with built-in utilities.
• Notebooks & Cookbooks: Get started quickly with Colab notebooks, how-to guides, end-to-end examples, and a cheatsheet available in the official documentation.