roboflow/supervision ? reverse-engineered prompt

Build me a Python computer vision toolkit that makes it easy to use different vision models without rewriting the rest of my code.

I want users to be able to run object detection, classification, and segmentation results through one common detections format, then draw clean boxes, labels, masks, and other annotations on images or video frames. It should also help with practical video tasks like tracking objects, counting things in zones, and making annotated output videos.

Please include simple dataset tools so someone can load image datasets, split them into train and test sets, merge datasets, and convert between common formats like COCO, YOLO, and Pascal VOC. Keep the API friendly and low friction, with clear examples that show loading an image, running a model result through the toolkit, annotating it, and saving the result.

Make it feel like a reusable library, not a one off script. Include tests and enough docs or examples so a beginner can copy and adapt it.