pytorch-pointcloud

A PyTorch library for deep learning on point clouds: models, pretrained weights, datasets, transforms and inferers, behind one create_model factory in the spirit of timm.

Object classification pointnet2-ssg.modelnet40.xu-yan	Part segmentation pointnext-sm.shapenetpart.openpoints	Indoor segmentation / detection ptv3-base.scannet20.pointcept
Outdoor segmentation / detection spvcnn-119gmacs.semantickitti.mit-han-lab second.kitti.openpcdet	Large scale segmentation utonia-lp.scannet20.pointcept	Features extraction sonata-lp.scannet20.fair

🎉 Highlights

• 35 architectures for classification, part and semantic segmentation, 3D detection and self-supervised pretraining: PointNet, PointNet++, DGCNN, KPConv, RandLA-Net, PointNeXt, PointMLP, PointConv, PVCNN, Point Transformer V1/V2/V3, SpUNet, SPVCNN, OctFormer, SphereFormer, Sonata, Concerto, Utonia, Point-MAE, Point-BERT, PointGPT, PointMamba, VoteNet, 3DETR, PointPillars, SECOND, PointRCNN, VoxelNeXt, LION and more.
• 134 pretrained checkpoints, each carrying its benchmark metrics measured through this library with the reference protocol, and loaded with a single create_model(..., pretrained=True).
• Datasets with download and preprocessing: ModelNet40, ScanObjectNN, ShapeNetPart, S3DIS, ScanNet, SemanticKITTI, nuScenes, KITTI, SUN RGB-D, Paris-Lille-3D, Semantic3D and Toronto3D.
• Transforms as MONAI-style dict transforms with tensor-level functional equivalents, and inferers for the usual evaluation protocols (test-time augmentation, voxel partition, sliding window, potential sphere voting).
• Packed batches everywhere: a flat $(N, \ldots)$ tensor plus a $(N,)$ batch index, never padded tensors.
• Fully typed (mypy strict), Python 3.10 to 3.13, an optional Lightning integration.

📦 Installation

pip install torch-pointcloud

The CUDA extensions (PyG kernels, spconv, flash-attention, Mamba, ocnn, torchsparse) are optional and only needed by the architectures that use them. See the Installation page for the exact install command.

🚀 Quickstart

import torch
import torch_pointcloud as tp

model = tp.create_model(
    "pointnext-sm.scanobjectnn.openpoints",
    task="classification",
    pretrained=True,
).eval()

pos = torch.randn(2048, 3)  # (N, 3) coordinates
x = torch.cat([pos, pos[:, 1:2] - pos[:, 1].min()], dim=1)  # (N, 4) features: xyz + height
batch = torch.zeros(2048, dtype=torch.long)  # (N,) batch index

with torch.no_grad():
    logits = model(x, pos, batch)  # (1, 15)

Every checkpoint ships the transform that turns a raw point cloud into what the network expects:

model, info = tp.create_model("ptv3-base.scannet20.pointcept", task="segmentation", pretrained=True, return_info=True)
info["transform"]  # the preprocessing pipeline of that checkpoint
info["weights"]["metrics"]  # {"mIoU": 76.29}

tp.list_models("pointnext*")  # every registered PointNeXt config
tp.list_models(task="detection", pretrained=True)  # all detection checkpoints

The examples directory for hands-on usage (benchmarks and training recipes).

📚 Documentation

The documentation covers installation, a get-started guide, the model zoo, datasets, transforms, tutorials and the full API reference.

📝 Citation

If you find this project useful, please consider citing:

@article{pytorch-pointcloud,
  title={PyTorch PointCloud},
  author={Arthur Dujardin},
  journal={GitHub},
  year={2026}
}

📄 License

Apache 2.0. See LICENSE.