LiDAR Training Data for Robotics Perception

It is the preparation of labeled point-cloud examples for robotics perception models. Labels describe objects, obstacles, workspace regions, and their changes across recorded scenes.

Yes. Use cuboids and other spatial shapes for object geometry, or point segmentation for class-labeled point regions. Define the label boundaries and expected model output before selecting a tool.

Yes. A scene can contain point-cloud sensors and camera images with calibration and poses. Annotators can inspect multiple views while keeping annotations tied to the recorded 3D scene.

Use persistent 3D object tracks, keyframes, interpolation, and visibility ranges. Camera-guided cuboid tracking is available when a suitable calibrated camera sequence is present; generated results still need review.

Define reusable classes, nested properties, and clear boundary and visibility rules. Use representative examples, required-property checks, and contextual review to resolve differences across recording sessions.

Select a cuboid and use Find Similar in a calibrated camera view to find similar objects. Fit candidate detections to the LiDAR points, review their classes and geometry, and accept the labels you need.

Upload supported PCD, PLY, and BIN point clouds through file, ZIP scene, or folder workflows. Use a scene manifest for frame sequences, sensor definitions, camera calibration, and poses where required.

Combine expert review with standalone-scene Consensus and approved Quality Gate benchmarks. Inspect geometry and properties in context, assign issues, and review corrections before creating an annotation release.

Export LiDAR annotations in Unitlab Unified Export Format (UUEF) with supported 3D geometry, tracks, scene metadata, calibration, and point-label arrays. Validate the exported labels against the coordinate and class conventions expected by your training pipeline.