3D point cloud annotation for lidar perception models Cuboids and point labels on every scan, consistent across the whole sequence
3D point cloud labeling for lidar and depth sensors Damco labels the scans your perception model learns to see from
How our lidar annotation process works From sample sequences to labeled scans your model can train on
- Week 1Agree the guideClasses, cuboid rules, and how to treat sparse and distant points.
- Week 2Run a pilotYou review a pilot sequence, and we fix the rules where labels disagreed.
- From week 3Annotate at volumeAgents propose cuboids, and annotators adjust them across frames.
- Every batchReview and sign offA reviewer replays sequences in 3D before the batch owner signs.
When to outsource 3D point cloud annotation And when 2D images or synthetic data will do
A good fit when
- You collect lidar for driving, delivery robots, or drones
- Your cuboids must stay consistent across long sequences
- You fuse lidar with camera images and need both labeled
Look elsewhere when
- Your model only uses camera images
- Synthetic scenes already cover the cases you need
- You have a few scans to label once
Lidar point cloud annotation in detail The techniques, the sensors, and how we keep sequences consistent
Why 3D point cloud labeling is hard
A lidar scan is a cloud of points with no color, and gaps where nothing reflected. A distant pedestrian may be a dozen points, and labeling it takes care and time.
Deep learning models for lidar need a great many labeled scans. That volume is why most teams outsource 3D point cloud annotation.
3D point annotation for all lidars
We label point clouds from spinning and solid-state lidars, and depth data from other sensors. Every point gets a class, so the object stands out from the ground, the walls, and the trees.
Object detection with 3D bounding boxes
A 3D bounding box, or cuboid, gives an object's location, size, and heading. Across a sequence it also shows speed, pitch, and direction, so the model learns how objects move.
Point cloud semantic segmentation
Segmentation assigns every point to a class such as road, sidewalk, curb, car, or tree. It lets the model tell kinds of lanes and surfaces apart in three dimensions.
Lane detection and object classification
We label lane lines and classify objects to fit the model's use case. Agents pre-label each frame, and annotators correct the result and keep each object's ID across the sequence.
3D point cloud annotation by industry
Lidar annotation serves autonomous vehicles, delivery robots, warehouses, agriculture, construction sites, and mapping. The guide changes with each one, because a forklift and a cyclist need different rules.
Questions about 3D point cloud annotation What teams ask before they outsource lidar labeling
What is 3D point cloud annotation?
It labels the points a lidar or depth sensor captures, so a model learns to recognize objects in 3D. The labels are usually cuboids around objects, a class for each point, or both.
Can you annotate lidar and camera data together?
Yes. We label the point cloud and the matching camera frames with the same objects, so each object keeps one ID across both. You send us the calibration between the sensors.
Which formats do you take and deliver?
We take point clouds as PCD, PLY, LAS, or BIN files. We deliver labels in KITTI format, as JSON, or in the format your tool reads.
How is 3D point cloud annotation priced?
We price per frame or per cuboid, depending on how dense your scenes are. The price comes from a pilot on your own sequences.
Send us a sample sequence and your classes
We'll reply with how we'd label it and the price per frame.