LiDAR & Point Cloud Annotation

3D perception datasets for autonomous systems — cuboids with correct orientation, point-level segmentation, and annotations kept consistent across LiDAR, RADAR, and camera.

Three-dimensional annotation is a different discipline

A 2D box has four degrees of freedom. A 3D cuboid has at least seven — position, extent, and heading — and heading is where most 3D datasets quietly go wrong. An object whose dimensions are right but whose yaw is ten degrees out still looks plausible on inspection, and it teaches a planner the wrong thing about where that vehicle is going.

Point clouds also get sparser with distance. A vehicle at fifteen metres may carry several hundred returns; the same vehicle at eighty may carry a dozen. Annotating the far object demands inference from partial evidence, and whether annotators are permitted to make that inference — and how it is recorded — has to be decided deliberately rather than left to individual judgement.

What we annotate in 3D

Our point cloud and sensor work covers:

  • 3D bounding boxes (cuboids) — position, dimensions, and heading, with a consistent rule for objects whose extent is only partially observed.
  • Point cloud semantic segmentation — per-point class assignment for road surface, vegetation, structures, and dynamic objects.
  • 3D object tracking — persistent identities across sequential sweeps.
  • RADAR annotation — labelling of returns and tracks, including cross-referencing against LiDAR and camera.
  • Multi-sensor fusion annotation — a single labelling decision applied consistently across every sensor that observed the object.
  • Lane and drivable-surface annotation — road geometry in three dimensions rather than projected into the image plane.

Sensor fusion and cross-modal consistency

When the same scene is captured by LiDAR, RADAR, and one or more cameras, the annotation has to agree across all of them. If an object is present in the point cloud but missing from the camera labels, a fusion model receives a contradiction rather than a supervision signal — and it will learn to distrust whichever sensor disagrees more often.

We annotate fused scenes as a single decision propagated across modalities, and check the cross-modal projection explicitly: an object labelled in 3D must land in the right place when projected into each camera frame. Calibration drift shows up clearly under that check, which is often the first time a team learns their extrinsics have moved.

Quality assurance in three dimensions

Every delivery passes through a layered review rather than a single pass. Annotators work to a written specification agreed before production starts; a second reviewer checks a defined proportion of each batch; and a final quality gate samples the batch independently against the same specification. Disagreements are not silently overwritten — they are resolved against the guideline, and where the guideline turns out to be ambiguous, the guideline is amended and the affected items are reworked.

That last step matters more than it sounds. Most systematic annotation error is not carelessness; it is a specification that failed to anticipate a real case. Treating every disagreement as a possible guideline defect is what keeps error rates from plateauing partway through a project.

Reviewing 3D annotation from a single viewpoint is unreliable — a cuboid that looks correct from above can be badly wrong in elevation. Our review passes inspect from multiple viewpoints and against the projected camera view, and we track orientation error separately from dimensional error, since the two have very different consequences for downstream planning.

Where this data is used

Point cloud annotation underpins perception for autonomous driving and ADAS, and increasingly for warehouse and field robotics, where the same geometry problems appear at a different scale. It also supports aerial and geospatial work, where LiDAR is used for terrain and structure modelling rather than for dynamic obstacle avoidance.

The annotation requirements differ substantially between those cases. A driving dataset lives or dies on dynamic object heading; a survey dataset generally does not care about heading at all and cares enormously about ground-versus-structure separation.

Data handling

Client data is encrypted in transit and at rest, access is limited to the specialists assigned to your project, and every member of the annotation workforce works under a signed confidentiality agreement. Where a dataset carries additional handling constraints, those are agreed in writing before any data is transferred.

Common questions

How do you annotate objects that return very few LiDAR points?

Sparse distant objects are handled by an explicit rule agreed before production: whether annotators may infer full extent from partial returns, and how such objects are flagged. They are marked so your team can weight or exclude them during training rather than discovering the uncertainty after the fact.

Can you annotate LiDAR and camera data together?

Yes. Fused scenes are annotated as one decision applied across every sensor that observed the object, and we verify that 3D annotations project correctly into each camera frame. That projection check also surfaces calibration drift.

Do you support RADAR as well as LiDAR?

Yes. RADAR returns and tracks are labelled and cross-referenced against LiDAR and camera observations of the same scene, which is what makes the data usable for sensor fusion rather than for RADAR-only models.

Which point cloud formats do you work with?

We work with the common point cloud and sequence formats used in autonomous driving datasets and export annotations to match your training pipeline. Format specifics are confirmed during project setup.

Discuss your lidar & point cloud project

Tell us what you are building and what your data has to support. We will come back with a specification, a pilot scope, and a realistic timeline.

Start a conversation