Supervision

Supervision sits between a vision model and the application around it. Detection, segmentation and classification models from Ultralytics, Transformers, MMDetection or Roboflow Inference all convert into one Detections object, and some libraries return it directly. From there the annotators draw the result, and they are meant to be composed rather than used one at a time.

The dataset side loads, splits, merges and saves in COCO, YOLO and Pascal VOC, including conversion between them, and images load on demand rather than all at once. It installs with pip on Python 3.10 or newer, and conda, mamba and source builds are documented.