Cuboid (3D bounding box)
A cuboid, also called a 3D bounding box, is an annotation type that encloses an object in three dimensions, capturing its position, size, and orientation in space, most often within a LiDAR point cloud.
In this article
Building visual or physical AI?
Let's talk.
What is a cuboid (3D bounding box)?
A cuboid is the 3D extension of a bounding box: instead of a rectangle on the image plane, it is a box in space defined by its center position, its dimensions along three axes, and its rotation. A common parameterization is 7 degrees of freedom (x, y, z, length, width, height, yaw), extended to 9 when pitch and roll matter. Cuboids are usually drawn over LiDAR point clouds, often fused with camera views so annotators can see the object from multiple angles. Each cuboid carries a class and can be tracked across frames.
Key takeaways
- A cuboid adds depth and orientation that a 2D box cannot represent.
- It is the core annotation type for autonomous driving and robotics perception.
- Orientation accuracy, not just overlap, is what makes a cuboid useful, because heading drives motion prediction.
Types and parameterization
- 7-DoF vs 9-DoF: yaw only, or full orientation with pitch and roll.
- Amodal vs visible: the full object extent even where occluded, or only the observed part.
- Single-frame vs tracked: a cuboid per frame, or one identity followed across a sequence.
How it works
Annotators place and orient the box in a 3D view, often snapping it to the point cluster and checking it against fused camera images.
Cuboid vs related labels
| Term | What it captures | When to use |
|---|---|---|
| Bounding box | 2D rectangle: location and extent on the image | Image and video objects |
| Cuboid | 3D box: location, size, and orientation in space | LiDAR and 3D scenes |
| Point cloud segmentation | Per-point class labels | Dense 3D scene understanding |
Why it matters
Cuboids give a perception model the spatial reasoning to know where an object is and which way it is heading, the difference between predicting a car's path and missing it. Standard axis-aligned IoU underweights orientation, so a cuboid can score a high IoU while pointing the wrong way, a real failure mode in autonomous-vehicle perception. That is why 3D tasks use rotated or 3D IoU, and why heading error is often reported separately from overlap.
Frequently asked questions
What is the difference between a cuboid and a bounding box?
A bounding box is 2D, a cuboid adds depth and orientation in 3D.
What data are cuboids drawn on?
Usually LiDAR point clouds, often fused with camera images.
Why does cuboid orientation matter?
Because heading drives motion prediction, and standard IoU can miss orientation errors.
Related terms
Learn more
See 3D detections and point clouds in a real self-driving workflow in Watching a Self-Driving Car Think.
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