51.2K ·GitHub
What is this?
Supervision is a Python library with reusable computer vision building blocks: detection/segmentation/classification result handling, annotators, trackers, and dataset utilities that plug into any model.
Supervision is a Python library with reusable computer vision building blocks: detection/segmentation/classification result handling, annotators, trackers, and dataset utilities that plug into any model.Is it a framework or a library?It's a model-agnostic library, not a framework or a specific model. It provides connectors to plug in results from libraries like Ultralytics, Transformers, MMDetection, Inference, or RF-DETR.
It's a model-agnostic library, not a framework or a specific model. It provides connectors to plug in results from libraries like Ultralytics, Transformers, MMDetection, Inference, or RF-DETR.How do I add it to a project?Install it with pip in a Python 3.10+ environment: pip install supervision. Conda, mamba, and source installs are covered in the guide.
Install it with pip in a Python 3.10+ environment: pip install supervision. Conda, mamba, and source installs are covered in the guide.Do I need a Roboflow API key?Only if you use the Roboflow Inference connector for running models; using other connectors like Ultralytics or Transformers does not require one.
Only if you use the Roboflow Inference connector for running models; using other connectors like Ultralytics or Transformers does not require one.Does it include its own detection models?No, Supervision doesn't ship models itself — you bring a model from a supported library and Supervision wraps the output into its sv.Detections format for annotation and analysis.
No, Supervision doesn't ship models itself — you bring a model from a supported library and Supervision wraps the output into its sv.Detections format for annotation and analysis.What can the annotators do?They let you compose highly customizable visualizations (boxes, masks, labels, traces, zones, etc.) on top of your model's detections to fit your specific use case.
They let you compose highly customizable visualizations (boxes, masks, labels, traces, zones, etc.) on top of your model's detections to fit your specific use case.Does it help with datasets too?Yes, it includes utilities to load, split, merge, convert, and save datasets in common formats like COCO and Pascal VOC.
Yes, it includes utilities to load, split, merge, convert, and save datasets in common formats like COCO and Pascal VOC.What kind of tutorials are available?Roboflow publishes end-to-end tutorials like speed estimation and vehicle tracking with YOLO and ByteTrack, and dwell-time/zone analysis for retail or traffic scenarios, built on top of Supervision.
Roboflow publishes end-to-end tutorials like speed estimation and vehicle tracking with YOLO and ByteTrack, and dwell-time/zone analysis for retail or traffic scenarios, built on top of Supervision.Which model libraries does it integrate with?It has ready connectors for Ultralytics, Hugging Face Transformers, MMDetection, Roboflow Inference, and models like RF-DETR that return sv.Detections directly.
It has ready connectors for Ultralytics, Hugging Face Transformers, MMDetection, Roboflow Inference, and models like RF-DETR that return sv.Detections directly.What are similar projects?X-AnyLabeling, deep-learning-for-image-processing, ultralytics, cvat, detr, mmdetection, yolov5, detectron2
X-AnyLabeling, deep-learning-for-image-processing, ultralytics, cvat, detr, mmdetection, yolov5, detectron2Which topics is it in?instance-segmentation, annotation, tracking, yolo, classification, metrics, low-code, object-detection, tensorflow, video-processing, image-processing, computer-vision
instance-segmentation, annotation, tracking, yolo, classification, metrics, low-code, object-detection, tensorflow, video-processing, image-processing, computer-vision