Object Detection

Object detection is a computer vision task that both labels objects in an image and locates them with bounding boxes.

Unlike image classification, which gives one label to the whole image, detection handles many objects at different places and sizes. Each detection is a class label, a box (x, y, width, height), and a confidence score. Two-stage detectors such as Faster R-CNN first propose regions likely to contain objects, then classify and refine each proposal. That is accurate and slow.

Single-stage detectors such as YOLO (You Only Look Once) and SSD predict boxes and classes in one pass, trading some accuracy for real-time speed. DETR (Detection Transformer) treats detection as set prediction with transformers and drops hand-designed pieces such as non-maximum suppression.

Uses include autonomous vehicles detecting pedestrians, vehicles, and signs; security systems spotting people and objects; industrial inspection finding defects; medical imaging locating tumors; and retail tracking inventory. COCO is the standard benchmark: precision and recall across 80 object categories.

Deployment still has to balance accuracy, speed, and behavior under messy lighting, occlusion, and weather.

A box is four numbers plus a class and a score. Faster R-CNN spends compute on region proposals, then a second stage. YOLO and SSD skip proposals and predict on a grid, which is why they hit real-time video. DETR outputs a set of detections with a transformer decoder and learns to avoid duplicates without non-maximum suppression.

COCO’s 80 classes and its precision-recall curves are how papers compare. A detector that is accurate in daylight can still fail in rain, glare, or heavy occlusion, which is the deployment tradeoff. YOLO (2016) framed detection as one pass over the image. Boxes and class scores come out together.