A desktop tool for 3D bounding boxes that skips the bloat

What it does

labelCloud is a desktop tool for annotating 3D object detection and 6D pose estimation datasets. You load a point cloud and draw oriented bounding boxes using either a two-click “picking” mode or a four-vertex “spanning” mode that locks depth layers to prevent accidental drift. Finished labels can be exported as centroids with Euler angles, vertex lists, or KITTI-compatible formats.

The interesting bit

Instead of wrestling with unconstrained 3D gizmos, the spanning mode lets you define a box by clicking four corners in sequence while the tool locks the width and height layers so you stay on the correct plane. There is also a semantic segmentation mode, but it works by assigning every point inside a bounding box to a class—using 3D boxes as coarse stencils rather than freehand brushes.

Key highlights

• Two labeling modes: picking for quick placement with mouse-wheel rotation, and spanning for precise vertex-by-vertex dimensioning.
• Imports colored and colorless point clouds including *.pcd, *.ply, *.xyz, and KITTI *.bin.
• Exports to centroid_rel, centroid_abs, vertices, and KITTI formats; you can subclass BaseLabelFormat to roll your own exporter.
• Defaults to z-axis-only rotation to keep boxes upright, with an optional 9-DoF mode for arbitrary orientations.
• Bounding-box-based semantic segmentation writes per-point class indices as *.bin files.

Caveats

• KITTI export requires calibration files, so it is not a zero-config option for raw point clouds.

Verdict

Use it if you want a fast, offline annotator for LiDAR or photogrammetry data and prefer keyboard shortcuts to cloud pipelines. Look elsewhere if you need multi-user web collaboration or point-level segmentation precision.