Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Nesti-Net: Normal Estimation for Unstructured 3D Point Clouds using Convolutional Neural Networks

Created by Yizhak (Itzik) Ben-Shabat, Michael Lindenbaum, and Anath Fischer from Technion, I.I.T.



This is the code for estimating normal vectors for unstructured 3D point clouds using Nesti-Net. It allows to train, test and evaluate our different normal estimation models. We provide the option to train a model or use a pretrained model. Please follow the installation instructions below.

Here is a short YouTube video providing a brief overview of the methods.


We propose a normal estimation method for unstructured 3D point clouds. This method, called Nesti-Net, builds on a new local point cloud representation which consists of multi-scale point statistics (MuPS), estimated on a local coarse Gaussian grid. This representation is a suitable input to a CNN architecture. The normals are estimated using a mixture-of-experts (MoE) architecture, which relies on a data-driven approach for selecting the optimal scale around each point and encourages sub-network specialization. Interesting insights into the network's resource distribution are provided. The scale prediction significantly improves robustness to different noise levels, point density variations and different levels of detail. We achieve state-of-the-art results on a benchmark synthetic dataset and present qualitative results on real scanned scenes.


If you find our work useful in your research, please cite our CVPR paper:

	author = {Ben-Shabat, Yizhak and Lindenbaum, Michael and Fischer, Anath},
	title = {Nesti-Net: Normal Estimation for Unstructured 3D Point Clouds Using Convolutional Neural Networks},
	booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
	month = {June},
	year = {2019}


  title={Nesti-Net: Normal Estimation for Unstructured 3D Point Clouds using Convolutional Neural Networks},
  author={Ben-Shabat, Yizhak and Lindenbaum, Michael and Fischer, Anath},
  journal={arXiv preprint arXiv:1812.00709},

Additionally, if you found this relevant to your work, you may also find our latest DeepFit paper and repo relevant.


Install Tensorflow and scikit-learn. You will also need torch, and torchvision for the PCPNet data loader.

The code was tested with Python 2.7, TensorFlow 1.12, torch 0.4.1, torchvision 0.2.1, CUDA 9.2.148, and cuDNN 7201 on Ubuntu 16.04.

Download the PCPNet data from this link and place it in the data directory. Alternatively, download the data used to train and test Nesti-Net from this link and place it in the data directory.

Download all trained models from this link and place them in the models directory. alternatively, dowload just the mixture of experts model (Nesti-Net) from this link


This repository allows to train a Nesti-Net multi-scale mixture of experts network for normal estimation. Simply run

This repository allows to train additional three vatiations: Single-scale / multi-scale models can be trained by running Multi-scale with switching - i.e. estimating the noise and switching between small scale network and large scale network (note that for this you will need a .txt file specifying the noise for each point cloud). It can be trained by running


To test Nesti-Net run and input the desired model log directory.

In order to test on your own data, place your point cloud directory in the data directory. Make sure that your point cloud directory includes a test set .txt file, that lists the files you wish to include in the test.

For the other models run the corresponding file.

Testing the different models will generate a results directory inside the trained model log directory. The results will be saved as separate .normals files containing the estimated normals.


To compute the RMS error, PGP5 or PGP10 evaluation metrics run for the desired results directory. The evaluation results will be saved in a summary directory within the results directory.


See LICENSE file.


Normal estimation for unstructured 3D point clouds using convolutional neural networks and a Mixture of Experts scale predictor.




No releases published


No packages published