Skip to main content
Zubnet AI › Learn › Wiki › YOLO
Models

YOLO

Also known as: You Only Look Once
A family of fast object detection models that find and classify all objects in an image in a single pass, the standard choice for real-time vision.

Why it matters

YOLO models are the workhorses of applied computer vision. They count products on shelves, spot defects on production lines, track players in sports broadcasts, measure traffic, guide drones and robots, and monitor wildlife. They're small enough to run on a phone, a camera or a cheap embedded board, and they can be fine-tuned on a custom dataset in an afternoon.

In an era of giant multimodal models, YOLO is a reminder that a small specialized model is often the right tool. It's orders of magnitude cheaper per image than a vision-language model, and far faster.

Deep Dive

YOLO stands for "You Only Look Once". Before it, the best object detectors worked in two stages: first propose regions that might contain something, then classify each region. That was accurate but slow. In 2015 Joseph Redmon and colleagues showed that a single neural network could look at the whole image once and directly predict every bounding box and its class. The first YOLO ran at 45 frames per second, fast enough for live video, at a time when accurate detectors managed only a handful.

Since then, "YOLO" has become less a single model than a lineage: more than a dozen versions from different teams, all built around the idea of fast one-stage detection.

How It Works

A YOLO model is a convolutional network with three parts: a backbone that extracts features, a neck that mixes features from different scales so that both small and large objects are visible, and a head that predicts boxes, confidence scores and classes for locations across the image.

The raw output contains many overlapping boxes for the same object. A post-processing step called non-maximum suppression (NMS) keeps the best one and removes the duplicates. Some recent versions are trained to output one box per object and skip NMS entirely, which simplifies deployment.

A Tangled Family Tree

Redmon wrote the first three versions, then left computer vision research in 2020, citing concerns about military and surveillance uses. The name lived on without an owner:

  • YOLOv4 (2020) came from Alexey Bochkovskiy and colleagues.
  • YOLOv5 (2020), YOLOv8 (2023) and YOLO11 (2024) came from the company Ultralytics, whose easy-to-use Python package made YOLO popular far beyond researchers.
  • YOLOv6, v7, v9, v10 and later versions came from various academic and industry groups.

The version numbers don't form a clean sequence of improvements, and a higher number isn't automatically better for your task. Modern YOLO packages also go beyond detection: the same family handles instance segmentation, pose estimation, rotated boxes and tracking.

Sizes and Metrics

Each version comes in several sizes, usually nano, small, medium, large and extra-large. Nano models run in real time on a phone, while larger ones are more accurate but need a GPU. Accuracy is reported as mAP (mean average precision) on the COCO dataset, which has 80 everyday object classes. For your own project, speed and accuracy on your data and your hardware matter more than COCO scores.

Training Your Own

Pretrained YOLO models know only COCO's 80 classes. For anything else, you fine-tune: collect a few hundred to a few thousand images, label them with bounding boxes and train for an hour or so on one GPU. Labeling is the real work, and tools built on Segment Anything or on vision-language models can speed it up considerably. For deployment, models are typically exported to ONNX, TensorRT, Core ML or similar formats.

Licensing and Alternatives

Watch the license. Ultralytics models are released under AGPL-3.0, which requires you to open-source software that uses them unless you buy a commercial license. Permissively licensed alternatives include YOLOX and transformer-based real-time detectors such as RT-DETR and RF-DETR, which now match or beat YOLO's accuracy.

Open-vocabulary detectors such as YOLO-World and Grounding DINO find objects described in text without any training, at some cost in speed and accuracy. Vision-language models can detect objects too, and they're useful for prototyping and labeling, but for high-volume or real-time work a small trained detector remains far more practical.

← All Terms
ESC