YOLO (You Only Look Once) is a family of one-stage architectures for object detection whose fundamental contribution was reformulating the task as a single regression solved in one pass of the network. Previous two-stage approaches first proposed candidate regions and then classified them one by one, which was computationally demanding. YOLO instead divides the image into a grid, and each cell simultaneously predicts the coordinates of several bounding boxes, their confidence scores and a probability distribution over classes. Because the model sees the whole image at once, it has global context available and therefore produces fewer false alarms on the background. At the cost of slightly lower accuracy on small and densely packed objects, it achieves a speed that enables real-time detection, including running on embedded devices. Development across versions brought anchor boxes and their later abandonment, multi-scale predictions at different resolutions, better augmentation, and architectural changes to heads and backbone. It is precisely this speed-to-accuracy ratio that makes YOLO the most widespread detector in industry – from quality control on a production line to traffic analysis and security monitoring.
Older methods worked like a person who takes a magnifying glass and systematically walks over a photograph patch by patch asking “is there a car here? And here? And here?” That is accurate, but it takes time. YOLO looks at the photograph in a single glance, the way you walk into a room and immediately know that two people are standing by the window and a laptop is lying on the table. No walking around required. The price of that speed is that you may not notice the ant in the corner – but when you have to decide twenty-five times a second because you are watching a camera feed, it is decidedly the better choice.