YOLO Object Detection: Real-Time Object Recognition with AI

Technical writer

Infrastructure

06.11.2025

17 min read

Imagine you are driving a car and in a split second you notice: a pedestrian on the left, a traffic light ahead, and a “yield” sign on the side. The brain instantly processes the image, recognizes what is where, and makes a decision.

Computers have learned to do this too. This is called object detection, a task in which you not only need to see what is in an image (for example, a dog), but also understand exactly where it is located. Neural networks are required for this. And one of the fastest and most popular ones is YOLO, or “You Only Look Once.” Now let’s break down what it does and why developers around the world love it.

What YOLO Object Detection Does

There is a simple task: to understand that there is a cat in a photo. Many neural networks can do this: we upload an image, and the model tells us, “Yes, there is a cat here.” This is called object recognition, or classification. All it does is assign a label to the image. No coordinates, no context. Just “cat, 87% confidence.”

Now let’s complicate things. We need not only to understand that there is a cat in the photo, but also to show exactly where it is sitting. And not one, but three cats. And not on a clean background, but among furniture, people, and toys. This requires a different task: YOLO object detection.

Here’s the difference:

Recognition (classification): one label for the entire image.
Detection: bounding boxes and labels inside the image: here’s the cat, here’s the ball, here’s the table.

There is also segmentation: when you need to color each pixel in the image and precisely outline the object's shape. But that’s a different story.

Object detection is like working with a group photo: you need to find yourself, your friends, and also mark where each person is standing. Not just “Natalie is in the frame,” but “Natalie is right there, between the plant and the cake.”

YOLO does exactly that: it searches, finds, and shows where and what is located in an image. And it does not do it step by step, but in one glance—more on that in the next section.

How YOLO Works: Explained Simply

YOLO stands for You Only Look Once, and that’s the whole idea. YOLO looks at the image once, as a whole, without cutting out pieces and scanning around like other algorithms do. This approach is called YOLO detection—fast analysis of the entire scene in a single pass. All it needs is one overall look to understand what is in the image and where exactly.

How Does Recognition Work?

Imagine the image is divided into a grid. Each cell is responsible for its own part of the picture, as if we placed an Excel table over the photo. This is how a YOLO object detection algorithm delegates responsibility to each cell.

An image of a girl on a bicycle overlaid with a 8×9 grid: an example of how YOLO labels an image.

Each cell then:

tries to determine whether there is an object (or part of an object) inside it,
predicts the coordinates of the bounding box (where exactly it is),
and indicates which class the object belongs to, for example, “car,” “person,” or “dog.”

If the center of an object falls into a cell, that cell is responsible for it. YOLO does not complicate things: each object has one responsible cell.

To better outline objects, YOLO predicts several bounding boxes for each cell, different in size and shape. After this, an important step begins: removing the excess.

What if the Neural Network Sees the Same Object Twice?

YOLO predicts several bounding boxes for each cell. For example, a bicycle might be outlined by three boxes with different confidence levels. To avoid chaos, a special filter is used: Non-Maximum Suppression (NMS). This is a mandatory step in YOLO detection that helps keep only the necessary boxes.

It works like this:

It compares all boxes claiming the same object.
Keeps only the one with the highest confidence.
Deletes the rest if they overlap too much.

As a result, we end up with one box per object, without duplicates.

What Do We Get?

YOLO outputs:

a list of objects: “car,” “bicycle,” “person”;
bounding box coordinates showing where they are located;
and the confidence level for each prediction: how sure the network is that it got it right.

An example of YOLO in action: the bicycle in the photo is outlined and labeled with its class and confidence score, and the image is divided into a 6×6 grid.

And all of this—in a single pass. No stitching, iteration, or sequential steps. Just: “look → predict everything at once.”

Why YOLO is Fast and What the “One Glance” Feature Means

Most neural networks that recognize objects work like this: first, find where an object might be, and then check what it is.

This is like searching for your keys by checking: under the table, then in the drawer, then behind the sofa. Slow, but careful.

YOLO works differently. It looks at the entire image at once and immediately says what is in it, where it is located, and how confident it is.

Imagine you walk into a room and instantly notice a cat on the left, a coat on the chair, and socks on the floor. The brain does not inspect each corner one by one; it sees the whole scene at once. YOLO does the same, just using a neural network.

Why this is fast:

YOLO is one large neural network. It does not split the work into stages like other algorithms do. No “candidate search” stage, then “verification.” Everything happens in one pass.
The image is split into a grid. Each cell analyzes whether there is an object in it. And if there is, it predicts what it is and where it is.
Fewer operations = higher speed. YOLO doesn’t run the image through dozens of models. That’s why it can run even on weak hardware, from drones to surveillance cameras.
Ideal for real-time. While other models are still thinking, YOLO has already shown the result. It is used where speed is critical: in drones, games, AR apps, smart cameras.

YOLO sacrifices some accuracy for speed. But for most tasks this is not critical. For example, if you are monitoring safety in a parking lot, you don’t need a perfectly outlined silhouette of a car. You need YOLO to quickly notice it and point out where it is.

That’s why YOLO is often chosen when speed is more important than millimeter precision. It’s not the best detective, but an excellent first responder.

How to Understand Whether a Neural Network Works Well

Let’s say the neural network found a bicycle in a photo. But how well did it do this? Maybe the box covers only half the wheel? Or maybe it confused a bicycle with a motorcycle?

To understand how accurate a neural network is, special metrics are used. There are several of them, and they all help answer the question: how well do predictions match reality? When training a YOLO model, these parameters are important—they affect the final accuracy.

IoU: How Accurately the Location Was Predicted

The most popular metric is IoU (Intersection over Union).

Imagine: there is a real box (human annotation) and a predicted box (from the neural network). If they almost match, great.

How IoU is calculated:

First, the area where the boxes overlap is calculated.
Then, the area they cover together.
We divide one by the other and get a value from 0 to 1. The closer to 1, the better.

Example:

Comment	IoU
Full match	1.0
Slightly off	0.6
Barely hit the object	0.2

An image of a bicycle with two overlapping rectangles: green for the human annotation and red for YOLO’s prediction. The rectangles partially overlap.

In practice, if IoU is above 0.5, the object is considered acceptably detected. If below, it’s an error.

Precision and Recall: Accuracy and Completeness

Two other important metrics are precision and recall.

Precision: out of all predicted objects, how many were correct.
Recall: out of all actual objects, how many were found.

Simple example:

The neural network found 5 objects. 4 of them are actually present; this is 80% precision. There were 6 objects in total. It found 4 out of 6—this is 66% recall.

High precision but low recall = the model is afraid to make mistakes and misses some objects.
High recall but low precision = the model is too bold and detects even what isn’t there.

AP and mAP: Averaged Evaluation

To avoid tracking many numbers manually, Average Precision (AP) is used. This is an averaged result between precision and recall across different thresholds.

AP is calculated for one class, for example, “bicycle”.

mAP (mean Average Precision) is the average AP across all classes: bicycles, people, buses, etc.

If YOLO shows mAP 0.6, this means it performs at 60% on average across all objects.

YOLO Architecture

From the outside, YOLO looks like a black box: you upload a photo and get a list of objects with bounding boxes. But inside, it’s quite logical. Let’s see how this neural network actually understands what’s in the image and where everything is located.

YOLO is a large neural network that looks at the entire image at once and immediately does three things: it identifies what is shown, where it is located, and how confident it is in each answer. It doesn’t process image regions step by step—it processes the whole scene in one go. That’s what makes it so fast.

To achieve this, it uses a special type of layer: convolutional layers. They act like filters that sequentially extract features. At first, they detect simple patterns—lines, corners, color transitions. Then they move on to more complex shapes: silhouettes, wheels, outlines of objects. In the final layers, the neural network begins to recognize familiar items: “this is a bicycle,” “this is a person”.

The main feature of YOLO is grid-based labeling. The image is divided into equal cells, and each cell becomes the “observer” of its own zone. If the center of an object falls within a cell, that cell takes responsibility: it predicts whether there’s an object, what type it is, and where exactly it’s located.

But to avoid confusion from multiple overlapping boxes (since YOLO often proposes several per object), a final-stage filter, Non-Maximum Suppression (NMS), is used. It keeps only the most confident bounding box and removes the rest if they’re too similar. The result is a clean, organized output: what’s in the image, where it is, and how confident YOLO is about each detection.

That’s YOLO from the inside: a fast, compact, and remarkably practical architecture, designed entirely for speed and efficiency.

How YOLO Evolved

Since YOLO’s debut in 2015, many versions have been released. Each new version isn’t just “a bit faster” or “a bit more accurate,” but a step forward—a new approach, new architectures, improved metrics. Below is a brief evolution of YOLO.

YOLOv1 (2015)

The version that started it all. YOLO introduced a revolutionary idea: instead of dividing the detection process into separate stages, do everything at once—detect and locate objects in a single pass. It worked fast, but struggled with small objects.

YOLOv2 (2016), also known as YOLO9000

Added anchor boxes—predefined bounding box shapes that helped detect objects of different sizes more accurately. Also introduced multi-scale training, enabling the model to better handle both large and small objects. The name “9000” refers to the number of classes YOLO could recognize.

YOLOv3 (2018)

A more powerful architecture using Darknet-53 instead of the previous network. Implemented a feature pyramid network (FPN) to detect objects at multiple scales. YOLOv3 became much more accurate, especially for small objects, while still operating in real time.

YOLOv4 (2020)

Developed by the community, without the original author’s involvement. Everything possible was improved: a new CSPNet backbone, optimized training, advanced data augmentation, smarter anchor boxes, DropBlock, and a “Bag of Freebies”—a set of methods to improve training speed and accuracy without increasing model size.

YOLOv5 (2020)

An open-source project by Ultralytics. It began as an unofficial continuation but quickly became the industry standard. It was easy to launch, simple to train, and worked efficiently on both CPU and GPU. Added SPP (Spatial Pyramid Pooling), improved anchor box handling, and introduced CIoU loss—a new loss function for more accurate learning.

YOLOv6 (2022)

Focused on device performance. Used a more compact network (EfficientNet-Lite) and improved detection in poor lighting and low-resolution conditions. Achieved a solid balance between accuracy and speed.

YOLOv7 (2022)

One of the fastest and most accurate models at the time. It supported up to 155 frames per second and handled small objects much better. Used focal loss to capture difficult objects and a new layer aggregation system for more efficient feature processing. Overall, it became one of the best real-time models available.

YOLOv8 (2023)

Introduced a user-friendly API, improved accuracy, and redesigned its architecture for modern PyTorch. Adapted for both CPU and GPU, supporting detection, segmentation, and classification tasks. YOLOv8 became the most beginner-friendly version and a solid foundation for advanced projects—capable of performing detection, segmentation, and classification simultaneously.

YOLOv9 (2024)

Designed with precision in mind. Developers improved how the neural network extracts features from images, enabling it to better capture fine details and handle complex scenes—for example, crowded photos with many people or objects. YOLOv9 became slightly slower than v8 but more accurate. It’s well-suited for tasks where precision is critical, such as medicine, manufacturing, or scientific research.

YOLOv10 (2024)

Introduced automatic anchor selection—no more manual tuning. Optimized for low-power devices, such as surveillance cameras or drones. Supports not only object detection but also segmentation (boundaries), human pose estimation, and object type recognition.

YOLOv11 (2024)

Maximum performance with minimal size. This version reduced model size by 22%, while increasing accuracy. YOLOv11 became faster, lighter, and smarter. It understands not only where an object is, but also the angle it’s oriented at, and can handle multiple task types—from detection to segmentation. Several versions were released—from the ultra-light YOLOv11n to the powerful production-ready YOLOv11x.

YOLOv12 (2025)

The most intelligent and accurate YOLO to date. This version completely reimagined the architecture: now the model doesn’t just “look” at an image but distributes attention across regions—like a human scanning a scene and focusing on key areas. This allows for more precise detection, especially in complex environments. YOLOv12 handles small details and crowded scenes better while maintaining speed. It’s slightly slower than the fastest versions, but its accuracy is higher. It’s suitable for everything: detection, segmentation, pose estimation, and oriented bounding boxes. The model is universal—it works on servers, cameras, drones, and smartphones. The lineup includes versions from the compact YOLO12n to the advanced YOLO12x.

Where YOLO Is Used in Real Life

YOLO isn’t confined to laboratories. It’s the neural network behind dozens of everyday technologies—often invisible, but critically important. That’s why how YOLO is used is a question not just for programmers, but for businesses as well.

In self-driving cars, YOLO serves as their “eyes.” While a human simply drives and looks around, the car must detect pedestrians, read road signs, distinguish cars, motorcycles, dogs, and cyclists—all in fractions of a second. YOLO enables this real-time perception without lengthy computations.

The same mechanisms power surveillance cameras. YOLO can distinguish a person from a moving shadow, detect abandoned objects, or alert when an unauthorized person enters a monitored area. This is crucial in airports, warehouses, and smart offices.

YOLO is also used in retail analytics—not at the checkout, but in behavioral tracking. It can monitor which shelves attract attention, how many people approach a display, which products are frequently picked up, and which are ignored. These insights become actionable analytics: retailers learn how shoppers move, what to rearrange, and what to remove.

In augmented reality, YOLO is indispensable. To “try on” glasses on your face or place a 3D object on a table via a phone camera, the system must first understand where that face or table is. YOLO performs this recognition quickly—even on mobile devices.

Drones with YOLO can recognize ground objects: people, animals, vehicles. This is used in search and rescue, military, and surveillance applications. It’s chosen not only for its accuracy but also for its compactness—YOLO can run even on limited hardware, which is vital for autonomous aerial systems. Such YOLO object detection helps rescuers locate targets faster.

Even in manufacturing, YOLO has applications. On an assembly line, it can detect product defects, count finished items, or check whether all components are in place. Robots with such systems work more safely: if a person enters the workspace, YOLO notices and triggers a stop command.

Everywhere there’s a camera and a need for fast recognition, YOLO can be used. It’s a simple, fast, and reliable system that, like an experienced worker, doesn’t argue or get distracted—it just does its job: sees and recognizes.

When YOLO Is Not the Best Choice

YOLO excels at speed, but like any technology, it has limitations.

The first weak point is small objects—for example, a distant person in a security camera or a bird in the sky. YOLO might miss them because it divides the image into large blocks, and tiny objects can “disappear” within the grid.

The second issue is crowded scenes—when many objects are close together, such as a crowd of people, a parking lot full of cars, or a busy market. YOLO can mix up boundaries, overlap boxes, or merge two objects into one.

The third is unstable conditions: poor lighting, motion blur, unusual angles, snow, or rain. YOLO can handle these to an extent, but not perfectly. If a scene is hard for a human to interpret, the neural network will struggle too.

Another limitation is fine-grained classification. YOLO isn’t specialized for subtle distinctions—for instance, differentiating cat breeds, car makes, or bird species. It’s great at distinguishing broad categories like “cat,” “dog,” or “car,” but not their nuances.

And finally, performance on weak hardware. YOLO is fast, but it’s still a neural network. On very low-powered devices—like microcontrollers or older smartphones—it might lag or fail to run. There are lightweight versions, but even they have limits.

This doesn’t mean YOLO is bad. It simply needs to be used with understanding. When speed is the priority, YOLO performs excellently. But if you need to analyze a scene in extreme detail, detect twenty objects with millimeter precision, and classify each one, you might need another model, even if it’s slower.

The Bottom Line

YOLO is like a person who quickly glances around and says, “Okay, there’s a car, a person, a bicycle.” No hesitation, no overthinking, no panic—just confident awareness.

It’s chosen for tasks that require real-time object recognition, such as drones, cameras, augmented reality, and autonomous vehicles. It delivers results almost instantly, and that’s what makes it so popular.

YOLO isn’t flawless—it can miss small objects or struggle in complex scenes. It doesn’t “think deeply” or provide lengthy explanations. But in a world where decisions must be made fast, it’s one of the best tools available.

If you’re just starting to explore computer vision, YOLO is a great way to understand how neural networks “see” the world. It shows that object recognition isn’t magic—it’s a structured process: divide, analyze, and outline.

And if you’re simply a user, not a programmer, now you know how self-checkout kiosks, surveillance systems, and AR try-ons work. Inside them, there might be a YOLO model doing one simple thing: looking. But it does it exceptionally well.

Infrastructure

06.11.2025

17 min read