A PyTorch framework for deep learning tasks on 3D point clouds like segmentation, classification, and registration.
Build, train, and benchmark point cloud neural networks using PyTorch Geometric and Facebook Hydra.
torch-points3d is an open-source Python framework designed for running advanced neural network architectures on 3D point cloud datasets. It combines PyTorch Geometric and Facebook Hydra to help technical teams train models for 3D object detection, segmentation, and registration. You should use it if you need a modular code structure for 3D spatial analysis.
The framework unifies popular 3D deep learning models—including PointNet++, KPConv, RandLA-Net, and MinkowskiEngine—into a single modular codebase. It organizes datasets and architectures by specific 3D tasks such as panoptic segmentation, classification, object detection, and registration. You can run training or evaluation pipelines using integrated support for benchmark datasets like ScanNet, S3DIS, and ShapeNet.
This tool suits machine learning engineers and 3D vision researchers analyzing LiDAR or photogrammetry scan data for engineering and architectural contexts. It is not designed for non-technical users or those looking for a graphical 3D modeling application. Teams without dedicated GPU hardware and Python development experience will find the setup challenging.
Organizes models and dataset loaders into clean subfolders for segmentation, classification, registration, object detection, and panoptic tasks.
Implements established 3D architectures including PointNet, PointNet++, KPConv, RandLA-Net, RSConv, and MinkowskiEngine.
Uses Facebook Hydra to manage experiment configurations, dataset parameters, and model hyperparameters reproducibly.
Includes built-in dataset handling for standard 3D benchmarks such as ScanNet, S3DIS, and ShapeNet.
torch-points3d is free to download and use directly from GitHub. There are no commercial hosted tiers provided by the maintainers, so your main cost will be the GPU compute infrastructure required for training.
877 open remote AI-training roles, updated November 16, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
These reviewed AI tools do the same job instantly — from our scored directory.
It supports 3D semantic segmentation, classification, part segmentation, 3D object detection, panoptic segmentation, and point cloud registration.
The framework includes dataset modules for ScanNet, S3DIS (in multiple sub-formats), and ShapeNet.
Yes, GPU hardware with CUDA 10 or higher is required for effective model training and sparse convolution operations.
See Zekai first in Google