LiDAR annotation: a complete guide to 3D point cloud labelling

3D cuboids, point cloud segmentation, multi-frame tracking and camera-LiDAR fusion, plus the mistakes to avoid and how to check quality.

LiDAR (light detection and ranging) measures distance with laser pulses and produces a point cloud: thousands to millions of 3D points describing the scene. Unlike a photo, every point has a real position in space, which makes LiDAR central to autonomous driving, robotics, drones and mapping. LiDAR annotation turns those points into labelled objects a model can learn from.

Main types of LiDAR annotation

TypeWhat it labelsUsed for
3D cuboids (3D bounding boxes)A box with position, size and heading around each objectObject detection for cars, trucks, pedestrians, cyclists
Semantic segmentationA class for every point (road, building, vegetation, pole)Scene understanding, HD maps, mapping and utilities
Instance segmentationSeparate IDs for each object of a classCounting and separating nearby objects
PolylinesLines for lanes, kerbs, power linesLane detection, corridor and asset mapping
Object trackingThe same ID for an object across framesMotion prediction and planning

Sensor fusion: LiDAR plus cameras

Most perception stacks combine LiDAR with cameras (and sometimes RADAR). In sensor fusion annotation, a cuboid drawn in 3D is projected onto each camera image so the labels agree across sensors. This needs accurate calibration (how each sensor is mounted) and time synchronisation. If the projected box does not sit on the object in the image, check calibration before blaming the annotator.

Common LiDAR annotation mistakes

  • Wrong heading. A box that fits the points but faces backwards. Always check yaw, especially on parked or slow objects.
  • Inconsistent box size. The same car should keep the same dimensions across frames, even when fewer points hit it.
  • Missed distant or occluded objects. Points get sparse with distance; guidelines should say how few points still count as an object.
  • Track ID switches. Two pedestrians crossing paths swap IDs. Review tracks, not just single frames.
  • Ground and noise points. Points from the road, rain or reflections included inside boxes or segmentation classes.

How to check LiDAR annotation quality

  • 3D IoU against reference boxes, plus a separate heading error check.
  • Dimension consistency for each tracked object over time.
  • Track continuity: no ID switches, no gaps when the object is visible.
  • Point-level accuracy per class for segmentation.
  • Projection check: 3D labels line up with the objects in camera images.

For the general metrics behind these checks, see annotation quality metrics.

Formats

Common outputs include the KITTI 3D label format and nuScenes-style JSON, as well as custom schemas that match your training code. Agree the coordinate frame (sensor, vehicle or world), units and heading convention before labelling starts; mismatched conventions are a frequent cause of "bad" labels that are actually fine.

Where LiDAR annotation is used

  • Autonomous vehicles and driver assistance
  • Warehouse and delivery robots
  • Drone and aerial LiDAR mapping
  • Utility corridors: poles, wires and vegetation clearance
  • Construction, mining and digital twins

LiDAR annotation at Annova

We label 3D cuboids, point cloud segmentation, LiDAR-camera fusion, RADAR and aerial LiDAR, with blind cross-verification on every batch. Our GIS background helps on aerial and utility LiDAR. See LiDAR & 3D point cloud annotation or request a free pilot.

Related services: LiDAR & 3D point cloud annotation · Autonomous vehicle annotation · Annova Annotation Platform

See our quality on your own data

Send a sample and your guidelines. We'll label a free pilot batch you can check before you commit.

Request a free pilot