PU-Net: Point Cloud Upsampling Network

PyTorch implementation of PU-Net. Official TF implementation: punet_tf. This repo is tested with PyTorch 1.2, cuda 10.0 and Python 3.6.

1. Installation

Follow Pointnet2.PyTorch to compile pointnet utils. Or run the following commands.

cd pointnet2
python setup.py install

You should install knn_cuda by running the following command or refering to KNN_CUDA

pip install --upgrade https://github.com/unlimblue/KNN_CUDA/releases/download/0.2/KNN_CUDA-0.2-py3-none-any.whl

2. Data Preparation

a. Prepare Patches

First, follow the official repo, download patches in HDF5 format from GoogleDrive and put it into ./datas/. Patches are splitted for training (3200) and testing (800). See ./datas/train_list.txt and ./datas/test_list.txt.

b. Prepare Datas for Visualization

Objects with 5k points for testing can be downloaded from the official repo, link. Put them into ./datas/test_data/our_collected_data/MC_5k.

c. Prepare Datas for NUC Calculation

The training and testing mesh files can be downloaded from GoogleDrive. Put test mesh files into ./datas/test_data/test_mesh.

The ./datas folder should be organized as follows:

PU-Net_pytorch
├── datas
│   ├── Patches_noHole_and_collected.h5
│   ├── test_list.txt
│   ├── train_list.txt
│   ├── test_data
│   │  │   ├── test_mesh
│   │  │   │   ├── *.off
│   │  │   ├── our_collected_data/MC_5k
│   │  │   │   ├── *.xyz

3. Train

Run the following commands for training.

mkdir logs
bash train_punet.sh

4. Evaluation (EMD and CD)

Run the following commands for evaluation.

python eval.py --gpu 0 --resume logs/punet_baseline/punet_epoch_99.pth

5. Visualization and Test (NUC)

Run the following commands to generate upsampled datas from full mesh objects with 5k points. Upsampled point clouds are saved in ./outputs/punet_baseline/*.ply. And the dumpped *.xyz files are used for NUC calculation.

mkdir outputs
bash test_punet.sh

NUC Calculation

1. install CGAL

2. run the following commands to compile cpp code

   cd nuc_utils
   mkdir build
   cd build
   cmake
   make
   cd ../..

3. run the following commands to calculate disk density, the results are saved in ./outputs/punet_baseline/.

   bash nuc_utils/evaluate_all.sh
   

4. run the following commands to calculate NUC

   python nuc_utils/calculate_nuc.py

Note that, the disk size (D) is 40 in default setting.

Performance

Please refer to this issue#1. I will update later.

Update

1. The auction matching is modified from PU-Net/code/tp_ops/emd. The number of points should be fewer than 4096 and better chosen as $2^K$ (e.g., 1024, 4096).
2. For the calculation of CD and EMD (evaluation), you should take the square root of the distance to get correct evaluation results.