Bounding box
A bounding box is an annotation type that marks an object's location in an image or video with a rectangle, usually stored as corner coordinates or as an origin point with width and height. It is the most common label for object detection.
In this article
Building visual or physical AI?
Let's talk.
What is a bounding box?
A bounding box is the simplest way to localize an object: a rectangle drawn tightly around it. It is typically axis-aligned and stored either as two corners (x_min, y_min, x_max, y_max) or as an origin plus size (x, y, width, height). Each box carries a class label, and a single image can hold many boxes. Boxes are the standard target for object detection, where the model learns to regress the rectangle and classify what is inside it.
Key takeaways
- The bounding box is the cheapest and most common annotation type, which makes it the default for object detection.
- It captures location and extent but not shape or orientation, so a tight box around an angled object still contains background.
- Box accuracy is scored with IoU, the overlap between the predicted and ground-truth rectangle.
Types and variations
- Axis-aligned vs rotated (oriented): a rotated box adds an angle to fit tilted objects.
- 2D vs 3D: the 3D analog is the cuboid, which adds depth and orientation.
- Tight vs amodal: whether the box hugs only the visible pixels or the object's full estimated extent.
How it works
Annotators draw the rectangle, often with snapping or model-assisted suggestions, and assign a class.
Before you train on boxes, compare them with the object's outline
A bounding box is the smallest axis-aligned rectangle around an object. It's quick to draw and enough for most detectors, but it also takes in whatever sits between the object's edges and the box's corners. The quiet problems:
- an aircraft parked at an angle, whose box is mostly tarmac
- boxes of neighboring planes that overlap, so it's unclear which wing belongs to which
- a box drawn to the fuselage on one plane and to the wingtips on the next
In Voxel51, RarePlanes has 5,000 satellite tiles from 112 airfields with aircraft boxes, outlines, and ten attributes per plane, so you can compare each box with the shape it stands for.
RarePlanes 5k
5,000 images
WorldView-3 satellite tiles of 112 airfields with aircraft boxes, outlines, and ten attributes per plane.
Try this: Find a plane parked at an angle and compare how much of its box the outline actually fills.
Open RarePlanes in Voxel51
Free with a work or .edu email.
Go deeper
Bounding box vs related labels
| Term | What it captures | When to use |
|---|---|---|
| Bounding box | 2D rectangle: location and extent | Object detection in images and video |
| Polygon or mask | Exact object shape at the pixel level | When shape matters |
| Cuboid | 3D box: location, size, and orientation | LiDAR and 3D scenes |
| Rotated (oriented) bounding box | A 2D box with an angle, to fit tilted objects | Aerial imagery, text, densely packed or angled objects |
| Keypoints | Specific points and the structure between them, not extent | Pose and landmark tasks |
Why it matters
Boxes are where most detection datasets and models start, because they are fast to draw and cheap to scale. A tight axis-aligned box around a diagonal or thin object, a cyclist's arm, a power line, can be mostly background, so the model trains on noisy positives. That is exactly when teams move to rotated boxes or segmentation. IoU thresholds also make box evaluation unforgiving, a model can locate an object yet still fail the 0.5 IoU bar on a loose box.
Frequently asked questions
How is a bounding box stored?
Either as corner coordinates (x_min, y_min, x_max, y_max) or as an origin point plus width and height.
What is the difference between a bounding box and a mask?
A box marks location with a rectangle, a mask marks exact shape at the pixel level.
How is bounding box accuracy measured?
With IoU, the overlap between the predicted and ground-truth box.
Related terms
Intersection over union (IoU), Cuboid / 3D bounding box, Polygon annotation, Object tracking, Annotation.
Learn more
See when a box stops being enough and a mask earns its cost in Preventing Your Vision AI Models From Failing, and how to tailor box styles to your data in Beyond Default Boxes.
Director of Growth
Jesse Mostipak is the Director of Growth at Voxel51, where the work is helping humans find and trust what the brand knows, and teaching the Google knowledge graph and the LLMs answering on their behalf to do the same. That question, how knowledge gets built inside a system, is one Jesse has been chasing for years. Earlier versions of it ran through a New York City high school science classroom, data science and machine learning, and developer relations at Kaggle, Posit (formerly RStudio), and Baseten. The answer doesn't change much depending on whether the learner is a teenager, a software engineer, or a knowledge graph. Jesse holds a Master's in Education from CUNY Hunter College. LinkedIn
See all articles by Jesse Mostipak
Last updated November 3, 2026
Building visual or physical AI?
Let's talk.
© 2026 Voxel51 All Rights Reserved