KPConv

Segments 3D point clouds using convolution defined directly on points rather than on a grid.

Field
Cross-domain
Method
Deep learning
Data it takes
point-cloud
License
MIT

KPConv applies convolution using kernel points positioned in space, so it operates on the raw cloud without voxelizing first. That avoids the resolution loss voxelization imposes on fine structure.

It predates the transformer-based approaches and remains a strong, well-understood baseline. Worth running alongside PTv3 when the question is which architecture suits a particular scan geometry.

Upstream project →

Catalog entry last checked 2026-08-19. All models →

Run KPConv on your data

The lab runs this on its own compute, in a container, with the inputs and parameters recorded alongside the result. Initial scoping conversations are free.

Contact the Lab