aboulch / Normals_houghcnn
normals_HoughCNN
Deep Learning for Robust Normal Estimation in Unstructured Point Clouds
Paper
Please acknowledge our the reference paper :
"Deep Learning for Robust Normal Estimation in Unstructured Point Clouds " by Alexandre Boulch and Renaud Marlet, Symposium of Geometry Processing 2016, Computer Graphics Forum
Dependencies
- Eigen and nanoflann (in the third_party_includes folder)
- TQDM
- Cython
- PyTorch
GPU support: NVIDIA GPU
The code has been tested on Ubuntu 18.04 with Anaconda.
Building the python library
cd path_to_repository/python
python setup.py install --home "."
It will build a library in the python folder of the repository.
Pretrained networks
Pretrained networks can be downloaded for 1, 3 and 5 scales (the models of the paper):
The models for previous versions (Lua Torch) can be downloaded here: 1 scale, 3 scales and 5 scales.
Usage
Once the library is built. You can use the test.py to test the estimation. The cube_100k.xyz file is located in the test directory.
Note: the input file must currently be at xyz format, it is possible to generate such file with Meshlab.
Note: number of scales has to be consistent with the used model (there are separate models for different scales).
Training from scratch
We provide the scripts for generating a training set and training a new model. The script train.py performs theses task. You can choose the scale number (1, 3 or 5) as in the paper.
Note: This is not the original training script from the related paper, but it should be similar. If you spot malfunctioning code or unexpected behavior, please contact the author.
License
The code is released under GPLv3 license. For commercial utilisation please contact the authors. The license is here.