Learn AI /Computer Vision and Language Technology /Bounding box
Bounding box
- Starter
- Computer Vision and Language Technology
- Technical readers
- Programme & data teams
Back to the milestone you came from
In five seconds
What it is
A rectangular marker that shows where an object or text region appears in an image, usually by its position, width and height.
In plain words
It is a quick rectangle around something important in an image so the system can say where it is.
An everyday example
At a distribution point, a supervisor reviews a photo of the storage area to estimate how many pallets of rice, water containers and medical kits are present. Instead of tracing the exact outline of each item stack, the supervisor draws a quick rectangle around each one on a printed copy. The rectangles make it easier to count, compare locations and discuss what needs moving. A very tall or oddly shaped stack may leave empty space inside the rectangle, but the marked area still shows roughly where that stack is.
A person drawing boxes on paper already knows what the items are and can use context that a model may miss. In real computer vision, the system must estimate both the object category and the rectangle from pixel patterns, and the box is stored as numbers, not hand-drawn marks. Also, some systems output many candidate boxes with confidence scores, which the paper example does not show.
Remember thisA bounding box records the approximate location of an object or text region as a rectangle. It does not trace the target's exact outline.
Understand it
Understand it
Why it matters
If you buy, assess or use vision software, you need to know what kind of location information it actually returns. A system that gives bounding boxes can support counting, tracking, routing, inventory checks, OCR pre-processing and quality control. But it may be too rough for tasks needing exact boundaries, such as measuring flood extent on the ground, isolating damaged building edges or separating overlapping objects. Knowing this term helps you ask better questions: How accurate are the boxes? What format are the coordinates in? How are overlapping boxes handled? Is a rectangle enough for the decision you need to make?
From image to bounding box
Show how a system turns an image into rectangular location data, and why that result is useful but approximate.
01
1. Input image
Start with a full image that may contain several objects or text regions.
Draw a simple scene in an SVG frame, such as a road with a bus, two people and a sign, with no markings yet.
02
2. Find a target
The system looks for patterns linked to a category, such as a bus or a text block.
Fade the background slightly and highlight one target object with a soft tint, without drawing a box yet.
03
3. Draw the rectangle
A rectangle is placed around the outer area of the target so its position and size can be recorded.
Overlay a clear rectangular stroke around the highlighted object; leave some empty space where the object's shape is irregular.
04
4. Output the result
The result is usually a label plus coordinates, such as left, top, width and height, sometimes with a confidence score.
Show a small side panel with example text like 'bus; x=120, y=60, w=180, h=90; confidence 0.93' linked by a line to the box.
A four-step diagram reads left to right. It starts with an image, then highlights a target object, then overlays a rectangle, then shows the rectangle converted into coordinates and a label. The reader should see that the box is a practical summary of location, not an exact outline.
On hover or tap, each step highlights one added element while keeping earlier elements visible, so the viewer can see what changes from step to step.
Describe this without the diagram
The diagram explains bounding boxes in four stages. First, there is an unmarked image containing several visible items. Second, one item is singled out as the target. Third, a rectangle appears around that item, showing that the system marks a rough outer area rather than the exact shape. Fourth, the rectangle is paired with output data: a class label, coordinates and possibly a confidence score. The main lesson is that a bounding box converts a visual location into a simple rectangular record that computers and people can use.
With reduced motion: Show the same four stages as static panels in one row. Each panel adds one element: the original image, the highlighted target, the rectangle, and the text output with coordinates.
A bounding box is a simple way to mark a region in an image. It is usually drawn as a rectangle around something important, such as a person, a vehicle, a building, or a line of printed text. The input is an image. The output is the rectangle's location, often written as coordinates.
A computer vision system may produce a bounding box after scanning the image for patterns. Coordinates are numbers that describe position. For example, a system might return the top-left corner of the rectangle plus its width and height, or it might return the left, top, right and bottom edges.
The box does not describe every pixel that belongs to the object. It only gives a rough enclosure. That is why it is often used for object detection, where the goal is to say what is present and where it is, quickly and in a standard format.
People often confuse a bounding box with image segmentation. Segmentation tries to mark the exact shape of the object, pixel by pixel. A bounding box uses a simpler representation, but is less precise, especially when the object has an irregular shape or overlaps with other things.
How “Drawing quick rectangles around supply items” maps onto the real thing
| In the story | In the AI system |
|---|---|
| quick rectangles around item stacks | bounding boxes around detected objects |
| photo of the storage area | input image |
| counting and locating stacks | using box coordinates for detection results |
| empty space inside a rectangle around an odd shape | the box encloses the object roughly rather than exactly |
A worked example
An ICT officer checks how a text-detection system marks three text regions in a synthetic event graphic before a side-event concept note goes for clearance.
Starting from: One image, 1280 by 720 pixels, containing an event title, a venue line and an access note.
- The system identifies three regions that match its learned patterns for text.
- For the event title, it places a rectangle with its top-left corner at x=340, y=210.
- It records the rectangle as 520 pixels wide and 80 pixels high.
- It returns the label "text" with those coordinates and a confidence score of 0.94.
Result: One detection can be recorded as text, x=340, y=210, width=520, height=80, confidence=0.94. Comparable records locate the venue line and access note.
In a United Nations workplace
At an inter-agency secretariat, an ICT officer reviews sample output for a document-accessibility demonstration before the concept note for a side event in the margins of the session goes into clearance. The tool has placed rectangles around headings and caption text in a mock event graphic.
Reading the coordinates, labels and confidence scores lets the officer verify what the tool has located and explain that each rectangle marks a text region rather than tracing the letters themselves.
Before you try thisThe test should use public or synthetic images. Draft concept notes, participant lists, email screenshots, or personal or security-sensitive material must not be uploaded to a tool not approved for that material, because an image can expose both visible content and embedded metadata.
Where you may hear it
- A vendor demonstration during a procurement meeting
- A review of text-recognition output before clearance
- An annotation briefing for a document-digitisation exercise
- A meeting of an ICT governance board
Avoid confusion
Avoid confusion
Often confused with
Image segmentation
Image segmentation divides an image into pixel-level regions; semantic segmentation labels them as roads, buildings or water.
Image classification
A system assigns one or more predefined labels to an image based on patterns in its pixels.
Pose estimation
A computer vision method that locates important body, face or hand points in an image or video and often links them into a skeleton-like layout.
Facial recognition
A computer vision method that compares face images or representations to estimate whether they show the same person.
Common misconceptions
Common misconception: A bounding box shows the exact shape of an object.
Correction: It usually does not. It encloses the object in a rectangle, which often includes background pixels and empty space. For exact shapes, segmentation is the more relevant term.
Common misconception: If a box appears, the system must have identified the object correctly.
Correction: Not necessarily. The rectangle can be well placed but attached to the wrong label, or the label can be right while the box is poorly sized. Location and classification quality both need checking.
Common misconception: Bounding boxes are only for photographs of physical objects.
Correction: They are also used for text regions in scanned documents, video frames, medical images and other visual data. The rectangle marks location, regardless of whether the target is a person, a paragraph or a parcel.
Quick check
Which statement best describes a bounding box?
A bounding box records rectangular location and size. The other options describe image classification, pose estimation and image segmentation.
Go deeper
Go deeper
The simple version
It is a quick rectangle around something important in an image so the system can say where it is.
The technical version
In computer vision, a bounding box is a rectangular localisation output that encloses a target instance or region of interest. It is commonly parameterised as x and y coordinates plus width and height, or as xmin, ymin, xmax and ymax, in either pixel units or normalised image coordinates. Detection systems predict these values, often together with a class label and a confidence score. Because the representation is axis-aligned in many systems, it is efficient and standardised, but it does not capture precise contours. Some applications use rotated rectangles or minimum bounding boxes, but those are variants rather than the default meaning in most deployed detection pipelines.
Reviewed external resources
EncyclopaediaMinimum bounding box algorithms (opens in a new tab)Wikipedia
This Wikipedia page gives background on minimum bounding box algorithms and suits readers who want mathematical or geometric context beyond everyday detection use.
Intermediatelink checked 2026-07-31
ArticleMachine Learning Glossary (opens in a new tab)Google
This glossary gives short plain-language definitions and suits readers who want a quick reference for related machine learning vocabulary.
Intermediatelink checked 2026-07-31
CourseMachine Learning Crash Course (opens in a new tab)Google
This course gives broader machine learning foundations and suits beginners who need surrounding concepts before applying terms used in vision systems.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Bounding box (opens in a new tab)YouTube search
This YouTube search results page helps readers find different video explanations of bounding boxes and suits people who learn well by watching, but results will vary.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Bounding box (opens in a new tab)Reddit search
This Reddit search results page helps readers see community questions and practical confusion points about bounding boxes and suits readers who want informal discussion, but quality varies.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
A bounding box records the approximate location of an object or text region as a rectangle. It does not trace the target's exact outline.