YOLO (You Only Look Once)

YOLO (“You Only Look Once”) is a single-stage object detector: one neural network that, in a single pass over the image, predicts at once the bounding boxes and classes of all the objects. Joseph Redmon and colleagues introduced it in 2015, framing detection as a direct regression problem—from pixels to coordinates and class—instead of the earlier approach of proposing regions and then classifying them. Its distinctive feature from the start is speed, which made it suitable for real-time video.

Detecting is not classifying

The concepts should be kept apart. Classification answers “what is there” and labels the whole image; detection answers “what and where,” locating and classifying each object at once; the bounding box is the rectangle around each detected object; and segmentation goes further, labeling pixel by pixel. YOLO does detection: it divides the image into a grid, and each cell predicts candidate boxes with their confidence and class. Modern versions are also multi-task, capable of segmentation, pose estimation or oriented detection.

A living family, not a single model

Today “YOLO” denotes not a model but an evolving family of detectors, so any figure on speed or accuracy requires stating the version. Redmon published versions 1 to 3 (between 2015 and 2018) and then left computer-vision research; from there the family fragmented across different teams. It has included YOLOv4 and v7 (Bochkovskiy, Wang and Liao), the versions from the company Ultralytics (v5, v8 and YOLO11), YOLOX (Megvii), YOLOv6 (Meituan), YOLOv9 (Academia Sinica) and YOLOv10 (Tsinghua University), which introduced detection without the non-maximum suppression stage. Among the most recent references are YOLOv13 (2025) and YOLO26, which Ultralytics released in early 2026 and describes as optimized for low-power devices.

Speed versus accuracy

YOLO was built to prioritize real time: by solving localization and classification in a single pass, it avoids the bottleneck of generating and evaluating thousands of candidate regions. Its historical price was lower accuracy on small objects or dense scenes. Two-stage detectors such as Faster R-CNN first propose regions and then classify them: more accurate in their day, but slower. Transformer-based detectors such as DETR frame detection without anchors or non-maximum suppression. Over the years the accuracy gap has narrowed and ideas have crossed between families: recent YOLO versions also drop non-maximum suppression.