Need help with PU-Net?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

251 Stars 62 Forks 39 Commits 16 Opened issues


PU-Net: Point Cloud Upsampling Network, CVPR, 2018 (

Services available


Need anything else?

Contributors list

# 74,298
37 commits

PU-Net: Point Cloud Upsampling Network

by Lequan Yu, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, Pheng-Ann Heng.


This repository is for our CVPR 2018 paper 'PU-Net: Point Cloud Upsampling Network'. The code is modified from PointNet++ and PointSetGeneration.


This repository is based on Tensorflow and the TF operators from PointNet++. Therefore, you need to install tensorflow and compile the TF operators.

For installing tensorflow, please follow the official instructions in here. The code is tested under TF1.3 (higher version should also work) and Python 2.7 on Ubuntu 16.04.

For compiling TF operators, please check
under each op subfolder in
folder. Note that you need to update
tensoflow include library
if necessary. You also need to remove
flag in g++ command in order to compile correctly if necessary.

To compile the operators in TF version >=1.4, you need to modify the compile scripts slightly.

First, find Tensorflow include and library paths.

    TF_INC=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_include())')
    TF_LIB=$(python -c 'import tensorflow as tf; print(tf.sysconfig.get_lib())')

Then, add flags of

-I$TF_INC/external/nsync/public -L$TF_LIB -ltensorflow_framework
to the
commands. You can refer


When running the code, if you have

undefined symbol: _ZTIN10tensorflow8OpKernelE
error, you need to compile the TF operators. If you have already added the
-I$TF_INC/external/nsync/public -L$TF_LIB -ltensorflow_framework
but still have
cannot find -ltensorflow_framework
error. Please use 'locate tensorflowframework ' to locate the tensorflowframework library and make sure this path is in


  1. Clone the repository:
   git clone
   cd PU-Net
  1. Compile the TF operators Follow the above information to compile the TF operators.

  2. Train the model: First, you need to download the training patches in HDF5 format from GoogleDrive and put it in folder

    . Then run:
    cd code
    python --phase train
  3. Evaluate the model: First, you need to download the pretrained model from GoogleDrive, extract it and put it in folder 'model'. Then run:

    cd code
    python --phase test --log_dir ../model/generator2_new6
    You will see the input and output results in the folder
  4. The training and testing mesh files can be downloaded from GoogleDrive.

Note: In this version, we treat the whole input point cloud as a single input. If the number of points in your input point cloud is big, you had better divide the input point cloud into patches and treat each patch as a single input. (see our following work EC-Net)

Evaluation code

We provide the code to calculate the metric NUC in the evaluation code folder. In order to use it, you need to install the CGAL library. Please refer this link to install this library. Then:

   cd evaluation_code
   cmake .
The second argument is the mesh, and the third one is the predicted points.

After running this program, the distances of each predicted point to the surface are written in
, and the density of each disk (ni/N) are written in ``.


If PU-Net is useful for your research, please consider citing:

     title={PU-Net: Point Cloud Upsampling Network},
     author={Yu, Lequan and Li, Xianzhi and Fu, Chi-Wing and Cohen-Or, Daniel and Heng, Pheng-Ann},
     booktitle = {Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
     year = {2018}



Please contact '[email protected]'

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.