Supervision for processing computer vision detections

How it works

Supervision provides common detection representations and tools for visualization and tracking. It can wrap results from several detector implementations.

What you need

Choose the adapter that maps your detector output into the library's representation. Keep the image dimensions and frame order available to verify the rendered result.

What to try first

Convert predictions from one detector and annotate a short video with track IDs. Check how the tracker handles a target leaving and returning to frame.

Include a dropped frame and a scene cut in the test video. The application should define what happens to existing track state across those boundaries.

Where to look for failures

Track IDs do not establish identity. Confirm coordinate conventions and detector confidence settings before comparing tracking behavior.

Related Franklin pages

Choose between text-guided boxes, prompted masks and downstream tracking. Preserve a few ambiguous images and videos so each stage can be evaluated independently.

• Train, evaluate and deploy supported YOLO tasks with a common Python and command-line interface. Ultralytics for training and exporting YOLO vision models
• Generate object masks from visual prompts and propagate them through video. SAM 2 for prompted image and video segmentation

For an overview of the types of predictions vision systems produce, see Franklin’s glossary. Computer vision

Project documentation

• Current project README
• README at the reviewed revision