YOLO

YOLO — short for "You Only Look Once" — is a style of object detector built for speed. Where the R-CNN family takes two passes (first propose regions, then classify them), YOLO does the whole job in a single glance at the image. It divides the picture into a grid and, for every grid cell at once, predicts what objects might be centered there and the boxes around them. One forward pass, all boxes out — hence the name.

That single-pass design is why YOLO can run in real time, processing dozens of video frames every second on ordinary hardware. This made it the go-to for anything live: dash cams, drones, sports tracking, factory lines, phone apps. The trade-off, candidly, is the classic speed-versus-accuracy tug-of-war. Early YOLO versions were noticeably less accurate than slower two-stage detectors, especially on small or clustered objects. Many later versions have narrowed that gap considerably, but the framing still holds: you choose YOLO when latency matters most.

A point worth keeping straight: "YOLO" is not a single fixed program but a long, evolving line of versions released by different groups over the years, each retuning the architecture. So a claim like "YOLO gets X accuracy" is meaningless without saying which version, on which dataset, at which speed. As always in vision, real-world reliability depends heavily on whether your scenes resemble the training data.