Until now the AI just said what a whole picture is. But a self-driving car needs to know where the people, cars, and signs are — with a box around each. That's object detection. Let's learn how we draw those boxes and grade them.
Same picture, two different questions. Classification gives the whole image one label. Detection finds every object and draws a box around each, with its own label. Tap the button to switch!
A detection is only good if its box actually lands on the object. We grade it with the overlap score: how much your box and the correct box share, divided by the total space they cover. Drag your cyan box onto the green target (drag the middle to move, the corner to resize).
IoU = area they share ÷ area they cover together. "Intersection over Union." If the boxes are identical, shared = total, so IoU = 1 (100%). If they don't touch, shared = 0, so IoU = 0. Detectors usually count a box as "correct" if IoU ≥ 50%.
Here's a scene with 3 objects. Drag each colored box onto the matching object. Your score is the average overlap across all three — get it to 70%+ to win, just like a real detector aims for.
This is YOLOS — a real object detector — running on your device. Point your camera at your desk, your room, your pets. It finds each object, draws a box, and labels it with a confidence score, live. (~10 MB one-time download; then cached and offline.)
Drawing and scoring boxes by hand shows the idea. Real detectors are trained networks:
✅ Shows how detection is graded. ⚠️ A person can't box millions of images.
Faster R-CNN - very accurate, a bit slower.
YOLO, SSD, RetinaNet (fast), DETR (transformer). ✅ Real-time. Judged by mAP.
You leveled up from what to where: object detection with boxes, and the overlap score (IoU) that grades every one.