Object Detection

Object detection is a computer vision task that simultaneously answers two questions about an image: what objects are present, and where are they located? Unlike simple image classification, which assigns a single label to an entire image, object detection produces bounding boxes paired with class labels for every object of interest in a scene. This dual requirement — classification plus localization — makes it significantly more complex and computationally demanding than either task alone.

Modern object detection systems are built on deep convolutional neural networks (CNNs) and generally fall into two architectural families. Two-stage detectors, pioneered by the R-CNN family, first propose candidate regions likely to contain objects and then classify each region independently, trading speed for accuracy. Single-stage detectors like YOLO (You Only Look Once) and SSD (Single Shot MultiBox Detector) predict bounding boxes and class probabilities directly from the full image in one forward pass, enabling real-time performance. More recently, transformer-based architectures such as DETR have reformulated detection as a set-prediction problem, eliminating hand-crafted components like anchor boxes and non-maximum suppression.

Training effective detectors requires large annotated datasets where every object instance is labeled with a bounding box and a category. Benchmarks such as PASCAL VOC and Microsoft COCO have standardized evaluation, using metrics like mean Average Precision (mAP) to measure how well predicted boxes overlap with ground-truth annotations across confidence thresholds. The availability of these datasets, combined with GPU-accelerated training, drove rapid accuracy improvements throughout the 2010s.

Object detection underpins a wide range of real-world applications: autonomous vehicles use it to track pedestrians, cyclists, and other cars; medical imaging systems flag tumors or anatomical landmarks; retail analytics count products on shelves; and security systems identify unauthorized individuals. As models have grown more accurate and efficient, deployment has expanded from cloud servers to edge devices and mobile hardware, making real-time detection feasible in resource-constrained environments.