YOLO
YOLO (You Only Look Once) is a family of real-time object detection models, originally introduced by Joseph Redmon in 2016, that detect and classify multiple objects in an image in a single forward pass through a neural network, making…
Definition
YOLO (You Only Look Once) is a family of real-time object detection models, originally introduced by Joseph Redmon in 2016, that detect and classify multiple objects in an image in a single forward pass through a neural network, making them fast enough for live video and embedded applications.
Overview
Before YOLO, many object detection systems worked in two stages: first proposing candidate regions of an image that might contain an object, then classifying each region separately — an approach that was accurate but too slow for real-time use. YOLO's core idea was to reframe object detection as a single regression problem: a single convolutional neural network looks at the whole image once and directly predicts bounding boxes and class probabilities for all objects simultaneously, which is where the name "You Only Look Once" comes from. The original YOLO divided an image into a grid, with each grid cell responsible for predicting bounding boxes and confidence scores for objects whose center falls within that cell. This single-pass design made YOLO dramatically faster than two-stage detectors of the time, trading some accuracy (particularly on small or overlapping objects) for real-time inference speed, which made it especially popular for video, robotics, and embedded applications where latency matters more than squeezing out the last percentage point of accuracy. YOLO has gone through many subsequent versions and maintainers since the original 2016 paper — YOLOv2 through YOLOv8 and beyond, developed by various teams and companies including Ultralytics — each iteration improving accuracy, adding features like instance segmentation and pose estimation, and refining the architecture with techniques such as anchor boxes, feature pyramid networks, and improved backbones like ResNet-style residual connections. Because of this long, community-driven iteration and the availability of open-source implementations — most commonly built and trained with PyTorch — YOLO became one of the most widely deployed object detection architectures in production computer vision systems, a topic covered hands-on in the PyTorch Deep Learning course.
Key Concepts
- Single-pass detection: predicts all bounding boxes and classes in one network forward pass
- Optimized for real-time inference speed, suitable for live video
- Divides images into a grid for localized bounding box prediction
- Widely used in embedded and edge devices due to its speed and small model variants
- Many community and vendor-maintained versions (YOLOv2 through YOLOv8+ and beyond)
- Extended in later versions to support segmentation and pose estimation