mit-han-lab/torchsparse

Health v2 · maintenance only

Adoption not part of the score

1472 stars · 192 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

TorchSparse is a high-performance PyTorch library for sparse convolution on 3D point clouds, with optimized GPU kernels for both training and inference. TorchSparse++ achieves significant speedups over prior sparse convolution libraries like MinkowskiEngine and SpConv.

Use cases

• run sparse convolutions on 3D point clouds fast
• train point cloud segmentation networks on GPU
• accelerate LiDAR perception models for autonomous driving
• speed up 3D object detection inference
• replace MinkowskiEngine or SpConv with a faster backend
• integrate sparse convolution into MMDetection3D or OpenPCDet
• 3D mesh and surface reconstruction from point clouds

When to choose

• you need state-of-the-art sparse convolution performance on NVIDIA GPUs
• you are building LiDAR or point cloud perception pipelines
• you want to plug into MMDetection3D or OpenPCDet
• you need both training and inference support for sparse CNNs

When to avoid

• you work with dense 2D images rather than sparse 3D data
• you need CPU-only or non-NVIDIA GPU support
• you want a turnkey application rather than a low-level library
• your project depends on a different sparse conv library's exact API

Facets

library · maturity active

machine-learning deep-learning gpu-computing image-processing computer-vision deep-learning autonomous-vehicles gpu-computing python sparse-convolution point-cloud pytorch 3d-deep-learning lidar gpu-acceleration autonomous-driving gpu linux cuda

Member repositories

For agents

markdown · JSON · MCP: product_card(name="mit-han-lab/torchsparse")