Model card for geotransformer.modelnet40.zheng-qin

A GeoTransformer point cloud registration model (geometric transformer for point cloud registration). Trained on ModelNet40.

Model Details

• Model Type: Point cloud registration
• Model Stats:
  • Params (M): 5.2
  • Input channels: 1
  • Features: 256
• Dataset: ModelNet40
• Metrics: RRE 1.597, RTE 0.019, RMSE 0.018
• Paper: Geometric Transformer for Fast and Robust Point Cloud Registration
• Converted from: qinzheng93/GeoTransformer (MIT)
• Library: torch-pointcloud

Install

pip install torch-pointcloud

Usage

import torch
import torch_pointcloud as tp
from torch_pointcloud.utils.data import collate

model, info = tp.create_model(
    "geotransformer.modelnet40.zheng-qin",
    pretrained=True,
    return_info=True,
)
model = model.eval()

# synthetic pair: a surface patch at the scale of the training scans, the source shifted off the reference
num_points = 8192
extent = torch.tensor([1, 1, 0.02])
sample = {
    "pos": torch.rand(num_points, 3) * extent + 0.1 * extent,
    "ref_pos": torch.rand(num_points, 3) * extent,
}
data = collate([sample], cat_keys=["ref_pos"])

with torch.no_grad():
    out = model(
        data.get("x"),
        data["pos"],
        data["batch"],
        data.get("ref_x"),
        data["ref_pos"],
        data["batch_ref_pos"],
    )
    registration = model.decode(out)  # transform (B, 4, 4) and the correspondences

Citation

@inproceedings{qin2022geotransformer,
  title   = {Geometric Transformer for Fast and Robust Point Cloud Registration},
  author  = {Zheng Qin and Hao Yu and Changjian Wang and Yulan Guo and Yuxing Peng and Kai Xu},
  booktitle = {CVPR},
  year    = {2022}
}

@inproceedings{wu2015modelnet,
  title   = {3D ShapeNets: A Deep Representation for Volumetric Shapes},
  author  = {Zhirong Wu and Shuran Song and Aditya Khosla and Fisher Yu and Linguang Zhang and Xiaoou Tang and Jianxiong Xiao},
  booktitle = {CVPR},
  year    = {2015}
}

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