What is a point cloud embedding?
A point cloud embedding does for 3D geometry what image embeddings do for pictures: it compresses a cloud of thousands or millions of 3D points into one vector whose neighbors are geometrically similar clouds. A LiDAR sweep of a highway on-ramp embeds near other on-ramps, and an unusual obstacle embeds far from everything.
This is harder than embedding images because point clouds are unordered, irregular, and sparse, which requires encoders designed for sets of points rather than pixel grids.
Key takeaways
- Point cloud embeddings make 3D geometry searchable by similarity.
- They require encoders built for unordered, irregular point sets.
- They enable find-similar, clustering, and outlier detection on LiDAR data.
How it works
Point cloud encoders either process raw points directly with permutation-invariant networks, voxelize the cloud into a 3D grid, or project it into views that 2D backbones can consume. The encoder pools its per-point or per-voxel features into a single vector, which is then indexed like any embedding. Sweeps, objects extracted from sweeps, or whole scenes can each be embedded at their own granularity.
Why it matters
LiDAR datasets are enormous and nearly impossible to review by eye, and point cloud embeddings give them the same navigability images now enjoy: retrieve similar sweeps, cluster recurring structures, and surface geometric outliers that may be sensor faults or genuinely novel scenes. That is foundational for curating the 3D side of autonomous driving and robotics data.
Frequently asked questions
Can image embedding models embed point clouds?
Not directly. Point clouds need encoders that handle unordered 3D points, though projecting a cloud to range or bird's-eye-view images lets 2D models approximate it.
What granularities can be embedded?
Whole sweeps, extracted objects, or aggregated scenes, with the right level depending on whether you search for situations or for individual structures.
Related terms
Go deeper