RegGeoNet: Learning Regular Representations for Large-Scale 3D Point Clouds

This is the official implementation of [RegGeoNet] (IJCV 2022), an unsupervised neural architecture to parameterize an unstructured 3D point cloud into a regular 2D image representation structure called deep geometry image (DeepGI), such that spatial coordinates of unordered 3D points are encoded in three-channel grid pixels.

<p align="center"> <img src="https://github.com/keeganhk/RegGeoNet/blob/master/imgs/toy_example.png" width="85%"> </p>

Setup

This code has been tested with Python 3.9, PyTorch 1.10.1, CUDA 11.1 and cuDNN 8.0.5 on Ubuntu 20.04.

• Install the differentiable computer vision library Kornia

pip install git+https://github.com/kornia/kornia

From Point Clouds to DeepGIs

Run para_pc.py under the scripts/para/ folder to convert a given 3D point cloud into its DeepGI representation.

Downstream Applications

Different downstream tasks can be performed directly on the generated DeepGIs, as an equivalent way of processing point cloud data. The pre-processed DeepGI-format datasets can be downloaded here (password: 5j86). Please put them under the data folder. Our pre-trained models are provided under the ckpt folder.

Citation

If you find our work useful in your research, please consider citing:

@article{zhang2022reggeonet,
  title={RegGeoNet: Learning Regular Representations for Large-Scale 3D Point Clouds},
  author={Zhang, Qijian and Hou, Junhui and Qian, Yue and Chan, Antoni B and Zhang, Juyong and He, Ying},
  journal={International Journal of Computer Vision},
  volume={130},
  number={12},
  pages={3100--3122},
  year={2022}
}