Does YOLO only classify a whole image?

No. Object detection predicts individual locations and associated classes or scores.

Does every YOLO model use the original grid and output format?

No. Later implementations can differ substantially, so version-specific documentation matters.

Can YOLO count moving objects by itself?

Frame detections can support counting, but reliable counts over time usually require additional tracking and event logic.

AI Glossary / Deep learning / Intermediate

You Only Look Once

Definition

You Only Look Once, usually abbreviated YOLO, is the name of an influential object-detection approach and a broad family of later detectors. The original model predicts bounding boxes and class-related scores from a whole image in one neural-network evaluation. It combines the questions “what is present?” and “where is it?” within a unified prediction system. A specific YOLO version and implementation must be identified because later designs differ substantially.

What it means

Image classification assigns a label to an image, while object detection identifies individual objects and their locations, usually with rectangular boxes. YOLO addresses detection by producing many candidate boxes and scores directly from image features. “Only look once” refers to the unified prediction approach, not a literal human-like glance or a guarantee that no preprocessing or postprocessing occurs. The name does not imply that every object will be found, that all classes are supported, or that the system tracks objects across video frames.

How it works

In the original design, the image is divided into a grid, and cells predict boxes and class information associated with object centers. Training compares predictions with labeled object locations and classes. At inference, candidate scores and locations are interpreted and filtered; the original pipeline uses non-maximum suppression to reduce duplicate detections. Later YOLO-family models can change the architecture, prediction representation, training objective, and postprocessing. Therefore, grid dimensions, confidence definitions, and anchor choices from one version should not be presented as universal properties of the entire family.

Three practical examples

Warehouse images: locate visible cartons or pallets before applying a separate inventory or inspection workflow.
Manufacturing inspection: detect known part types in images, then evaluate whether additional measurements are needed to assess defects.
Video analytics: detect objects in individual frames as an input to a separate tracking or counting system, with careful handling of repeated detections.

Think of it this way

Think of marking all likely objects on a photograph in one coordinated pass rather than treating each possible crop as a separate classification exercise. The analogy captures the unified prediction idea. Its limit is that the model still performs many numerical operations and may need filtering, and a box is a prediction rather than proof that an object is present.

Worked example

Suppose an illustrative detector produces two same-class boxes in a 100-by-100-pixel image. Box A runs from (10, 10) to (30, 30) with score 0.9. Box B runs from (12, 12) to (32, 32) with score 0.8. Each has area 400 under a continuous-coordinate area convention. Their overlap has width 18 and height 18, so intersection area is 324. Union area is 400 + 400 - 324 = 476. Their intersection over union, or IoU, is therefore about 0.6807.
Under a stated same-class non-maximum suppression policy with IoU threshold 0.5, A is retained first and B is suppressed as a likely duplicate. If the threshold were 0.7, B would not be removed by this pairwise comparison. This is a generic postprocessing example, not a claim that every YOLO version uses that policy. Suppression can reduce duplicate boxes but can also remove a legitimate nearby object in crowded scenes. The scores are illustrative ranking values; they should not automatically be read as calibrated probabilities of correctness.

Illustrative case study

A fictional warehouse team tests a YOLO-family detector on camera images of stacked cartons. Performance on clear daytime examples looks good, but partially hidden and very small cartons are often missed. The team evaluates by object size and lighting condition, reviews annotation consistency, and measures complete camera-to-result latency on its actual hardware. It separates detection from counting so repeated boxes across frames do not inflate inventory totals. The project report names the exact model and evaluation setup instead of quoting a generic YOLO speed claim.

Common misconceptions and mistakes

YOLO is not a single unchanged architecture, and a higher version label alone is not an accuracy guarantee. Bounding boxes are not pixel-precise segmentation masks. Detection confidence does not necessarily equal a calibrated probability, and a detector trained on fixed classes may mislabel unfamiliar objects. Frames per second depends on input size, hardware, batching, and which processing stages are timed. Finding an object in each frame is also different from maintaining its identity over time.

Business use

A YOLO-family detector can be useful when object locations are needed within a practical latency budget. Specify target classes, annotation rules, acceptable misses and false alarms, object-size distribution, and operating conditions. Compare models on representative held-out images and measure the full deployment pipeline. Inspect precision and recall at the threshold actually used, along with appropriate detection metrics such as average precision under stated IoU conventions. Include model version, weights, preprocessing, and runtime configuration in reproducibility records.

Check your understanding

Why is box B suppressed at IoU threshold 0.5 in the example?

Answer: It has a lower score than A and overlaps A with IoU about 0.6807, exceeding the stated same-class suppression threshold. This treats B as a likely duplicate under that policy, not as mathematical proof that only one real object exists.

Related terms

Bounding box
Computer vision
Convolutional neural network
Intersection over union
Precision
Recall
Object detection
non-maximum suppression
average precision
object tracking

Learn next

Learn bounding-box coordinates and IoU, then compare image classification, object detection, segmentation, and tracking. Read the original YOLO design to understand its motivation, followed by the documentation of the exact implementation being evaluated. Practice interpreting errors and latency measurements under clearly stated operating conditions.

Bounding box

Sources

Redmon, Divvala, Girshick, and Farhadi, You Only Look Once: Unified, Real-Time Object Detection: https://arxiv.org/abs/1506.02640 Original authors' YOLO project: https://pjreddie.com/darknet/yolo/ Original CVPR paper: https://openaccess.thecvf.com/content_cvpr_2016/papers/Redmon_You_Only_Look_CVPR_2016_paper.pdf

Previous: Yoked control design Next: Youden's J statistic