Skip to main content
Some models need more than a flat rectangle: vehicle pose, parcel dimensions, furniture orientation. The Cuboid tool draws a pseudo-3D box on a 2D image: a front face plus projected depth edges, giving downstream models the geometry a bounding box throws away.
Cuboids annotate perspective on flat images. For true 3D boxes on point clouds, that is the LiDAR editor’s territory with a dedicated project type.

Draw a cuboid

1

Arm the class and pick the tool

Press the class hotkey, then click the cuboid icon in the floating palette.
2

Two clicks for the front face

Click one corner of the object’s front face, then the opposite corner, exactly like the two corners of a bounding box.
3

One click for depth

Move toward the object’s receding direction; the preview extends the depth edges. Click to set the depth point.
4

Enter to finish

Enter completes the cuboid, and it joins Object Instances like any other annotation.
Orange cuboid annotation on a white sedan in the PixlHub editor showing front face and depth edges

A finished cuboid on a sedan: filled front face, wireframe depth edges following the car's orientation.

Getting the geometry right

  • Front face on the visible face. Put the rectangle on the surface facing the camera (a car’s front or rear); the depth edges then follow the body naturally.
  • Depth follows the vanishing direction. Drag depth along the object’s real-world axis, using lane lines, shelf edges or building lines as perspective references.
  • Edit like everything else. Select the cuboid and its handles expose both face corners and depth for adjustment; reclassify with a hotkey, delete with Del, undo per operation.
PixlHub editor with a cuboid annotation on a sedan in a busy street scene

The cuboid in scene context: perspective encoded that a flat box cannot express.

When cuboid beats box

Frequently asked questions

It is perspective-annotated 2D: a front face plus a projected depth, in image coordinates. Genuinely metric 3D comes from LiDAR projects, where annotation happens in a point cloud.
With their full geometry (face and depth points), alongside other annotation types; downstream code reads them as eight-point projections. See Exports.

Next steps

Bounding Box

The flat sibling, for when pose does not matter.

Image annotation editor

The full tour of tools and panels.