SpConv
lastmod 2026-11-10created 2026-03-312 edits
spconv is a convolution library for spatially sparse 3D data. When a LiDAR point cloud is voxelized, most voxels are empty; spconv implements sparse convolution and submanifold sparse convolution in CUDA, computing only the voxels that hold values, which greatly reduces memory use and computation. It is provided as a PyTorch extension and is widely used as a backbone in 3D object detection and point cloud segmentation models.
Spatially Sparse Convolution Library