Boxes follow what it sees, live.
Each frame, the model puts a box around everything it recognises, from people and cars to cups and phones, 80 kinds in all. The slider sets how sure it has to be before a box shows.
In 2024 I built a car detector with YOLOv5. YOLO makes many overlapping guesses for each object and needs a clean-up step, non-maximum suppression, to keep one box each. RF-DETR is a detection transformer: each of its 300 slots learns to claim a different object, so there is nothing to clean up.
Runs on your device. Camera frames are never uploaded. The two street photos are CC0 from Wikimedia Commons, by Wilfredor and Ingolfson; the glove and the living room are mine.