Fork of "Orienting Point Clouds with Dipole Propagation" by Metzer et al.

Modified solely to resolve package version conflicts for compatibility with our code. See the original README below for full details.

The following changes where applied: File: /dipole-normal-prop/utils.py

time.clock() -> time.perf_counter()
torch.arange(n)[mask] -> torch.arange(n, device=mask.device)[mask]
e, v = torch.symeig(cov, eigenvectors=True) -> e, v = torch.linalg.eigh(cov, UPLO='L')

File: /dipole-normal-prop/inference_utils.py

e, v = torch.symeig(cov, eigenvectors=True) -> e, v = torch.linalg.eigh(cov, UPLO='L')

File: /dipole-normal-prop/field_utils.py

min_index = np.array([curv[i][0] for i in range(len(all_patches))]) -> min_index = np.array([curv[i][0].cpu().numpy() for i in range(len(all_patches))])

Orienting Point Clouds with Dipole Propagation

Leveraging the power of neural networks and a robust and stable dipole propagation technique.

SIGGRAPH 2021 [Paper] [Project Page]

by Gal Metzer, Rana Hanocka, Denis Zorin, Raja Giryes, Daniele Panozzo, and Daniel Cohen-Or

Getting Started

Installation

• Clone this repo

Setup Conda Environment

To run the code, it is advised to create a new conda environment by running conda create -n dipole python=3.6
and activate it by running conda activate dipole

then install the required libraries below:

• PyTorch version 1.6.0 (might work with later versions as well)
  install from PyTorch getting started with your appropriate CUDATOOLKIT
• Pytorch Geometric
  may use the helper installation script ./install_pytorch_geometric.sh or from here.
  ** if using the helper script make sure to change the declared variables inside the script to match your installed versions **
• argparse pip install argparse
• Open3D pip install open3d (optional for faster normal estimation)

./install_cu101.sh is a full installation script for pytorch 1.6.0 and cudatoolkit 10.1.
./install_cu102.sh is a full installation script for pytorch 1.8.0 and cudatoolkit 10.2.
Make sure to change /path-to-anaconda3 in the installation script to the path on your machine

Running Examples

The code contains three main orientation scripts

• orient_pointcloud.py used to orient relatively small point clouds, taking into account all input points.
• orient_large.py used to orient relatively large point clouds, using a representative set of points for each patch.
• reference_orientation.py used to transfer the orientation from an oriented source point cloud to an un-oriented target point cloud.

The /demos folder contains demo scripts for running the code on point clouds from the /data folder.

Simple cd demos and run any of the scripts:

Visualizing Examples

The oriented point clouds can easialy be visualized by opening them with MeshLab.

Example Point Clouds

• fandisk.sh
• hand.sh (Table 1 - 17_42l)
• ok.sh (Table 1 - 09_41r)
• vase.sh
• galera.sh
• boxunion.sh
• flower.sh

Large Example Point Clouds

Example on a large point clouds

• lion.sh (Figure 17)
• alien.sh (Figure 18)

Reference orientation

Example of using the dipole field to transfer the orientation from an oriented point cloud to nearby points which are unoriented (e.g., generated using upsampling or consolidation).

• reference_orientation.sh calculated on an output of Self-Sampling
  which does not have normal information for the produced consolidated point cloud. (Figure 15)

Simple Point Propagation

Example of using dipole propagation individually per point i.e. without patching and network steps.

• ok_simple.sh (Same shape as Table 1 - 09_41r)

In this implantation we added an extra step to fix the global orientation such that the normals are not only consistent,
but also points outside the surface, which was not included in the original paper.

This technique relies on the fact that the dipole potential for a closed and correctly oriented shape is zero outside, and strictly positive or negative inside. Depending on whether the normals point inside or outside.
The dipole potential is measured on a square lattice around the oriented point cloud, and the sign of the average potential is used to determine whether all normals point outside.

Data

Data thanks to Self-Sampling NestiNet, Mano, threedscans, COSEG.

Citation

If you find this code useful, please consider citing the paper

@article{metzer2021orienting,
author = {Metzer, Gal and Hanocka, Rana and Zorin, Denis and Giryes, Raja and Panozzo, Daniele and Cohen-Or, Daniel},
title = {Orienting Point Clouds with Dipole Propagation},
year = {2021},
publisher = {Association for Computing Machinery},
volume = {40},
number = {4},
journal = {ACM Trans. Graph.},
}

Questions / Issues

If you have questions or issues running this code, please open an issue.