pytorch-pointcloud

A PyTorch library for deep learning on point clouds: models, pretrained weights, datasets, transforms and inferers, inspired by timm.

Check out the documentation at pytorch-pointcloud.org!

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	Feature extraction sonata-lp.scannet20.fair

Installation

pip install torch-pointcloud

Most models also need pyg-lib, and some a CUDA extension (spconv, flash-attn, ...), built for your torch and CUDA versions. The installation page writes the commands.

Quickstart

import torch
import torch_pointcloud as tp

# Requires pyg-lib
model = tp.create_model(
    "pointnext-sm.scanobjectnn-hardest.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)

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

import torch_pointcloud as tp

# Requires spconv and flash-attn
model, info = tp.create_model("ptv3-base.scannet20.pointcept", task="semantic-segmentation", pretrained=True, return_info=True)
info["transform"]  # the preprocessing pipeline of that checkpoint
info["weights"]["metrics"]  # {"mIoU": 77.40, "OA": 92.01}

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

See the examples directory for benchmarks and training recipes.

Why torch-pointcloud?

torch-pointcloud is a library of pretrained models that makes common layers and backbones easy to reuse. It does not replace research-first codebases such as Pointcept, OpenPCDet and MMDetection3D or OpenPoints but it complements them with a common interface that makes benchmarking and interoperability across architectures easier. It builds on PyG for the packed batch format and the neighbor search, and adds what PyG does not ship for point clouds: the models, their weights, the datasets and the transforms.

Documentation

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

Contributing

Contributions are welcome. See CONTRIBUTING.md for the development setup and the pull request checks.

Citation

If you find this project useful, please consider citing:

@software{dujardin2026pytorchpointcloud,
  author = {Dujardin, Arthur},
  title = {PyTorch PointCloud},
  year = {2026},
  url = {https://github.com/arthurdjn/pytorch-pointcloud},
  doi = {10.5281/zenodo.22159632},
  license = {Apache-2.0}
}

License

Apache 2.0. See LICENSE.

Pretrained weights and adapted code keep the license of their source, and some checkpoints are restricted to non-commercial use. Most were trained on research-only datasets, whose terms also apply to the weights. See THIRD_PARTY_NOTICES.md.