object detection

Object detection is the vision task of finding things in an image and saying both what each one is and where it is. The 'where' is usually drawn as a box — a rectangle, called a bounding box, snugly wrapped around each object — and the 'what' is a label like 'person', 'cup', or 'forklift' stuck to that box. So a single photo of a kitchen might come back marked with one box around the mug, another around the kettle, and a third around the cat on the counter, each tagged with its name.

This is different from just classifying a picture. Plain image classification only answers 'is there a cat in this photo, yes or no?' and assumes the picture is mostly one thing. Detection is harder because a real scene holds many objects at once, of different sizes, sometimes overlapping or half hidden, and the system has to pin down each one separately. For a robot this matters enormously: an arm cannot pick up a bottle it has only vaguely classified as 'present somewhere in the frame' — it needs the bottle's exact location to reach for it, and it needs to tell the bottle apart from the glass next to it.

Modern object detectors are almost always built on neural networks that learn from huge collections of labeled example images, with well-known families including YOLO, SSD, and Faster R-CNN. For every candidate location it considers, the network predicts a class label, a confidence score (how sure it is), and the four numbers that define the box. The output is a fast, frame-by-frame answer to 'what is in front of me and where,' which is the starting point for grasping, obstacle avoidance, counting, and tracking.