punet_yongchi/punet_pytorch

• PU-Net: Point Cloud Upsampling Network
• • 1. Dataset
  • 2. Install extensions including KNN-CUDA and Pointnet2
  • 3. Train or Test
  • 4. Evaluation
  • 5. Performance
  • 6. Acknowledgement

PU-Net: Point Cloud Upsampling Network

This is PyTorch implementation of PU-Net based on Official TF repo punet_tf. The code has been tested with Python 3.6.9, PyTorch 1.2.0 and CUDA 10.2 on a Tesla T4.

1. Dataset

We have provided official patched data in HDF5 format for trainning, which are stored in './datas/'. To output upsampled points, 20 objects with 5k points are put into ./datas/test_data/our_collected_data/MC_5k. For evaluation, test mesh files are downloaded into ./datas/test_data/test_mesh.

2. Install extensions including KNN-CUDA and Pointnet2

# KNN-CUDA offline installation
pip install KNN_CUDA-0.2-py3-none-any.whl

# Pointnet2
cd extensions/pointnet2
python setup.py install

3. Train or Test

You can switch 'mode' to train or test. The trained model will be stored in '.weights/punet' and the upsampled points for test will be stored in './results/punet'.

python main.py --mode train

python test.py --mode test

4. Evaluation

1. Firstly, you should install CGAL library by running

   apt-get update
   apt-get upgrade
   apt-get install libcgal-dev libcgal-demo
   

2. compile cpp code

   cd evaluation_code
   cmake .
   make
   

3. generate density file.

   cd /evaluation_code
   sh evaluate_all.sh
   

4. calculate NUC metric

   python nuc_utils/calculate_nuc.py
   

5. Performance

NUC with different p

0.002	0.004	0.006	0.008	0.010	0.012	0.015
0.17464873	0.14155353	0.12641019	0.11760626	0.11187568	0.11156724	0.11221573

6. Acknowledgement

PU-Net
punet_tf
punet_pytorch
knn-cuda
pointnet2