yolo

YOLO — You Only Look Once (Redmon et al., 2016) — reframed detection from a multi-step pipeline into a single regression problem solved in one forward pass, which is why it is the archetypal real-time detector. Where two-stage methods propose-then-classify, YOLO looks at the whole image once and directly outputs all boxes and class probabilities together. 'You only look once' is both the name and the design philosophy: no proposals, no re-cropping, no second pass.

The original idea: divide the image into an S×S grid (e.g. 7×7). Each grid cell is responsible for objects whose center falls in it, and predicts B bounding boxes (each with x, y, w, h and a confidence reflecting objectness × box quality) plus C class probabilities for the cell. The whole output is one tensor of shape S×S×(B·5 + C), produced by a single CNN and trained with one combined loss over localization, confidence, and classification. Seeing the global image at once gives YOLO strong context and very few background false positives.

YOLO's original trade-off was speed for precision: it ran at 45+ fps (Fast YOLO over 150 fps) but localized less accurately than Faster R-CNN, and struggled with small or clustered objects because each cell predicted few boxes and one class. Successive versions systematically closed that gap. v2 (YOLO9000) added anchor boxes, batch norm, and dimension clusters; v3 added multi-scale prediction (an FPN-like three-scale head) and a stronger backbone; v4/v5 industrialized the training tricks (mosaic augmentation, CIoU loss, PANet neck).

Modern YOLO (v6–v12 and the Ultralytics line) are anchor-free, use decoupled classification/regression heads, advanced label assignment (e.g. SimOTA/TAL), and distribution-focal box regression, reaching accuracy competitive with two-stage detectors while staying real-time. The brand has become an umbrella for fast single-stage detection rather than one fixed architecture; the constant is the philosophy — one network, one pass, dense predictions on a grid, NMS at the end (until very recent end-to-end NMS-free YOLO variants).