A 3D bounding box, also called a cuboid, is a labeling method that draws a box around an object in three dimensions, capturing its length, width, height, and orientation, rather than just marking it on a flat image. It's most often applied to LiDAR point cloud data, where each object needs a box that reflects its actual position and size in physical space.

‍

‍Why it matters

A flat 2D box can tell a model where an object appears on screen, but it can't tell the model how far away that object actually is or which direction it's facing. For a self-driving car, that's the difference between knowing there's a pedestrian somewhere ahead and knowing exactly how many meters away they are and which way they're walking. The extra dimension is what makes depth and motion prediction possible at all.

‍

‍How it's labeled

‍Annotators typically work inside a 3D viewer, rotating and placing a box around a point cluster from multiple angles to make sure it actually fits the object rather than just looking right from one flat view. Tools like CVAT support this kind of point cloud annotation directly. Because getting the orientation right takes more time than drawing a flat box, 3D annotation projects usually run slower and cost more per object than standard 2D bounding boxes.

‍

‍Where it's used

Autonomous vehicle systems depend on this heavily, since LiDAR sensors capture the world as a 3D point cloud and the model needs cuboids around every car, pedestrian, and obstacle to navigate safely. Robotics uses the same idea, a robot arm or mobile robot needs to understand the exact size and position of an object before it can pick it up or move around it. Drone mapping and warehouse inventory systems apply it too, tracking the volume and placement of objects in physical space.