Object detection

Classification says what is in an image; detection also says where — a bounding box plus a class plus a confidence, for every object at once.

How YOLO-style detectors work

• The image is processed once by a CNN ("You Only Look Once") — fast enough for a robot's live camera.
• The network proposes boxes everywhere, scores each (class + confidence), and near-duplicate boxes are merged.
• Output per frame: a list like ball (0.97) at [x, y, w, h].

From detection to behavior

A detection is just data — behavior comes from the loop around it:

1. Detect the target in the frame.
2. Compare its box position to the frame center → steering error.
3. Compare its box size to a target size → distance, so speed.
4. Repeat every frame — the perception → control loop.

This is exactly what the object detection practice builds, and why "follow me" robots feel alive.

Training your own detector

Same supervised recipe, richer labels: instead of a class per image you draw boxes per image. All the data lessons from supervised learning apply doubly — box quality is label quality.