Cameras capture flat pictures. LiDAR and depth sensors capture shape: millions of points, each with an exact position in space, that together describe the world in three dimensions. That collection of points is a point cloud, and it is the raw material for 3D perception in autonomous vehicles, robots, drones and digital twins. This guide explains what a point cloud is, how point cloud data is created and stored, how point clouds are used in machine learning and what point cloud annotation involves.
In This Article
Key Points
- A point cloud is a set of points in 3D space, each with X, Y and Z coordinates and often intensity or color.
- Most point clouds come from LiDAR, photogrammetry or depth cameras, and they are sparse, unordered and uneven in density.
- Machine learning on point clouds powers 3D detection, segmentation and mapping for vehicles, robots and drones.
- Point cloud annotation uses 3D cuboids, point-level segmentation and polylines, usually alongside camera images.
What Is a Point Cloud?
A point cloud is a collection of data points in three-dimensional space, where each point records a position (X, Y and Z coordinates) and often extra values such as intensity, color or time, together describing the shape of objects and surfaces. Seen on screen, a dense point cloud looks like a 3D photograph made of dots; zoomed in, it is simply a long list of coordinates.
Unlike an image, a point cloud has no fixed grid. Points are scattered wherever the sensor’s signal returned from a surface, so they are dense on nearby objects and sparse far away, and they contain nothing where a surface was hidden. Those properties are what make point clouds powerful for measuring the world, and what make them harder to process and label than images.
How Point Cloud Data Is Created
The most common source is LiDAR, which fires laser pulses and measures how long each takes to bounce back. The US Geological Survey describes raw lidar output as a point cloud of individual points reflected from everything on the surface, which it uses to build high-resolution elevation models. Vehicles and robots carry spinning or solid-state LiDAR that produces a new point cloud many times a second.
Point clouds also come from photogrammetry, which reconstructs 3D points from overlapping photos, and from depth cameras and structured-light sensors used indoors and in robotics. Airborne LiDAR on planes and drones captures terrain, buildings and vegetation over large areas, as our geospatial annotation work shows.
Point Cloud Formats and Attributes
Beyond position, points often carry intensity (how strongly the surface reflected the pulse), return number (useful for seeing through vegetation), color from a paired camera, and a timestamp. Data is stored in formats such as LAS and LAZ for survey and mapping work, PLY and PCD in robotics and research, and sensor-specific binary formats in autonomous driving. Programs that combine LiDAR with cameras also need calibration data so each point can be projected onto the matching image.
Point Clouds in Machine Learning
Point cloud machine learning trains models to understand 3D scenes directly from points. Common tasks include 3D object detection, which finds vehicles, people and obstacles with their size and heading; point cloud segmentation, which labels every point by class; tracking objects across sweeps; and mapping and localization, which match live scans to a stored map. Models either work on raw points, group them into small 3D cells called voxels, or project them into bird’s-eye or range images that image models can process.
All of these approaches learn from annotated examples, and the sparse, uneven nature of point clouds makes those examples harder to produce than image labels. A pedestrian 60 meters away may be a few dozen points; a car partly hidden behind a truck may be half missing. Training data has to cover those situations with consistent rules.
What Is Point Cloud Annotation?
Point cloud annotation is the labeling of 3D point data so models can learn from it. The main methods are 3D cuboids, which box each object with its position, dimensions and heading; point-level semantic segmentation, which assigns a class to every point; instance labels that separate individual objects; and 3D polylines for lanes, curbs and other linear features. Most autonomous programs also link 3D labels to 2D camera labels so the same object carries the same ID across sensors.
Good 3D point cloud annotation needs rules for heading direction, how far a cuboid extends around sparse or partly hidden objects, which points to include at object edges, and how to handle reflections and noise such as rain or dust. Our guides to cuboid yaw orientation and LiDAR and sensor fusion annotation go deeper on both.
Common Point Cloud Use Cases
- Autonomous vehicles: 3D detection and tracking of road users, and lane-level HD map features.
- Robotics and warehouses: obstacle avoidance, shelf and pallet detection, and grasp planning.
- Mapping and surveying: terrain, buildings and vegetation from airborne and mobile LiDAR.
- Infrastructure and construction: digital twins, progress tracking and inspection of bridges, power lines and rail.
How Annotera Supports Point Cloud Projects
Annotera labels LiDAR and other point cloud data with 3D cuboids, point-level segmentation, polylines and cross-sensor links to camera images, for driving, robotics, mapping and infrastructure programs. Work runs to written 3D rules with gold sets and three-layer QA, and starts with a free 48-hour pilot on your own sweeps. See our LiDAR annotation services for the full scope.
Frequently Asked Questions
What is a point cloud in simple terms?
It is a 3D picture made of dots. Each dot is a measured point on a surface with an exact position in space, usually captured by LiDAR or reconstructed from photos.
What is the difference between a point cloud and a 3D model?
A point cloud is raw measured points with no connections between them. A 3D model, such as a mesh, connects points into surfaces and is often built from a point cloud.
How are point clouds used in machine learning?
Models learn from labeled point clouds to detect and track 3D objects, segment scenes point by point and match scans to maps, mainly for autonomous vehicles, robots and mapping.
What is 3D point cloud annotation?
It is the labeling of point cloud data with 3D cuboids, point-level classes, instance IDs and polylines so models can learn 3D perception tasks.
Getting Started
Point clouds give AI systems a measured, three-dimensional view of the world. Turning them into training data takes clear 3D rules, consistent labels across sensors and QA built for sparse, uneven data.
Working with LiDAR or point cloud data? Talk to Annotera about a pilot on your own data.
