YOLO (You Only Look Once)
YOLO: You Only Look Once — Unified, Real-Time Object Detection · Also known as: You Only Look Once, YOLO detector, YOLOv1, single-shot detector, real-time object detection
YOLO (You Only Look Once) is a single-shot, end-to-end convolutional object detector introduced by Redmon, Divvala, Girshick, and Farhadi at CVPR 2016. It reframes object detection as a single regression problem — predicting bounding box coordinates and class probabilities directly from an image in one forward pass — achieving real-time detection speeds that prior two-stage methods such as R-CNN could not match. The original paper spawned a widely adopted family of successors (YOLOv2 through v11) that continues to dominate applied object detection benchmarks.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
YOLO and its successors are appropriate when detection speed is a hard constraint — live video surveillance, autonomous driving, drone imagery, robotic perception — or when a single network must localise and classify multiple object categories simultaneously. The method assumes objects of interest are large enough to be captured by the grid resolution; it performs less well on very small or densely clustered objects. A GPU is required for real-time throughput. For tasks where detection latency is not critical and maximum accuracy is paramount, two-stage detectors such as Faster R-CNN remain competitive.
Strengths & limitations
- Processes the entire image in a single forward pass, enabling real-time detection speeds (45 fps at CVPR 2016; modern versions are faster still).
- End-to-end training on detection loss means the model optimises for the full detection objective rather than a proxy task.
- Global image context is implicitly encoded, which reduces false positives from local texture patterns compared with sliding-window methods.
- Highly extensible: the YOLO architecture has been revised continuously (v2–v11), with anchor boxes, multi-scale prediction, attention modules, and transformer heads incorporated in later versions.
- Straightforward deployment: a single network with no separate proposal stage makes integration into embedded and edge systems practical.
- The fixed S×S grid imposes a hard limit on the number of objects detectable per cell; YOLOv1 can predict at most B objects per cell, causing misses when many small objects cluster in one region.
- Bounding box dimensions and aspect ratios are fully regressed, without the anchor-box priors used in later detectors, making it harder to fit unusual object shapes.
- Localisation accuracy is lower than two-stage detectors such as Faster R-CNN on standard benchmarks such as PASCAL VOC and COCO, particularly for small objects.
- Training requires a large annotated dataset with bounding-box labels; performance degrades substantially when labelled data are scarce.
Frequently asked
How does YOLO differ from Faster R-CNN?
Faster R-CNN uses a two-stage pipeline: a Region Proposal Network first generates candidate object regions, which are then classified and refined by a second network. YOLO collapses both stages into a single forward pass over a grid, trading some localisation accuracy for substantially higher throughput. On PASCAL VOC, YOLOv1 ran roughly 6× faster than Faster R-CNN but at lower mAP; later YOLO versions close much of this accuracy gap.
Which YOLO version should I use for a new project?
For most practical applications, a recent version (v8 or later) from an actively maintained repository is recommended, as these incorporate improved backbones, anchor-free heads, and data augmentation strategies that substantially outperform YOLOv1 in accuracy and robustness. The original YOLOv1 is primarily of historical and educational interest.
What dataset and annotation format does YOLO training require?
YOLO training requires images paired with bounding-box annotations in (class, x_centre, y_centre, width, height) format normalised to [0, 1]. Popular training sets include PASCAL VOC and MS COCO. Transfer learning from a COCO-pretrained checkpoint is standard practice when labelled data are limited.
How is detection performance measured?
The standard metric is mean Average Precision (mAP), computed as the area under the precision–recall curve averaged over object classes. The COCO benchmark reports mAP averaged across IoU thresholds from 0.50 to 0.95 (mAP@[0.5:0.95]), which is more demanding than the single-threshold PASCAL VOC metric (mAP@0.5). Inference speed is reported as frames per second (fps) on a specified GPU.
Sources
- Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016). You Only Look Once: Unified, Real-Time Object Detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 779–788. DOI: 10.1109/CVPR.2016.91 ↗
- Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. ISBN: 978-0-262-03561-3
How to cite this page
ScholarGate. (2026, June 3). YOLO: You Only Look Once — Unified, Real-Time Object Detection. ScholarGate. https://scholargate.app/en/deep-learning/yolo
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
Compare side by side →