object detection
/ OB-jekt dee-TEK-shun /
Object detection answers a richer question than classification: not just what is in the picture, but what and where, for everything at once. The output is a set of boxes drawn around each object, each box carrying a label and a confidence — "person, 0.88" here, "car, 0.95" there, "dog, 0.71" in the corner. A single street photo might come back with a dozen boxes. Where classification gives one verdict for the whole image, detection localizes and labels many things in parallel.
Think of it as combining two jobs: drawing a tight rectangle (a bounding box) around something, and naming what is inside. Early systems did this slowly by scanning the image with sliding windows; modern ones learn to propose likely regions and label them in one swift pass. The model also has to decide how many objects there are — a crowd scene and an empty desk demand very different numbers of boxes — which makes the task genuinely harder than picking one label.
This is the workhorse behind self-driving perception, retail shelf-scanning, security cameras, and photo apps that tag faces. But the honest caveats are real: detectors miss small, distant, or overlapping objects; they fire false alarms on shadows and reflections; and their accuracy drops sharply on objects, lighting, or camera angles unlike their training data. A detector that is excellent on daytime city streets can be unreliable at night or in heavy rain, which is exactly why safety-critical systems never trust it alone.
Run a detector on a photo of a crosswalk and it returns: a box around each of 4 pedestrians (person, 0.9+), a box around a cyclist (bicycle 0.82, person 0.79), and two cars. It missed the dog half-hidden behind a leg, and drew a spurious box on a mannequin in a shop window labeled person 0.55.
Many boxes, many labels, all at once — plus the typical misses and false alarms.
A bounding box is a crude shape: it is always an upright rectangle, so it includes background around tilted or irregular objects and cannot separate two things that overlap inside the same box. When you need the exact silhouette, you want segmentation, not detection.