humanoidsdata.com

Search

Search companies, datasets, articles, and glossary terms for humanoids and embodied AI.

← All glossary terms

Data & collection

Point cloud

A point cloud is a collection of points representing sampled locations in two- or three-dimensional space. Each point has coordinates in a stated reference frame and may also carry fields such as colour, intensity, timestamp, surface normal, semantic label, or sensor-specific metadata.

Also known as: 3D point cloud, point-cloud data, point clouds

Updated

A point is more than three numbers

Point clouds commonly come from LiDAR, structured-light or time-of-flight cameras, stereo reconstruction, multi-view geometry or simulation. Cartesian x, y and z coordinates locate a sample, but its meaning also depends on the sensor origin, reference frame, units and capture time.

The ROS PointCloud2 message represents a timestamped set of points with a coordinate frame and an extensible set of fields. A cloud may therefore contain colour, reflectance, ring index or other values alongside position. Consumers should inspect the schema instead of assuming that every point has the same layout.

Organised and unorganised clouds behave differently

The Point Cloud Library distinguishes organised and unorganised point clouds. An organised cloud retains a row-and-column structure like an image, which can preserve neighbourhood relationships from a depth camera. An unorganised cloud is a flat set of points, as is common after merging scans or filtering samples.

A depth image and a point cloud can encode closely related geometry, but they are not interchangeable. The depth image stores distance on an image grid and relies on camera calibration for back-projection. A point cloud stores coordinates that have already been expressed in a frame, potentially after projection, motion compensation or registration.

Dataset quality depends on geometry and provenance

Point count alone says little about usefulness. Density changes with range and viewpoint; reflective or transparent surfaces can create missing or spurious returns; moving objects can distort a scan; and downsampling can remove small structures important for grasping or foot placement.

A robotics dataset should document the generating sensor or simulator, fields and data types, units, coordinate frame, calibration, timestamps, invalid-point convention, filtering and registration steps. If several captures were merged, retain the poses and method used. Labels should specify whether they apply to individual points, objects, voxels or the source image.

Sources