PU-Flow: a Point Cloud Upsampling Network with Normalizing Flows

by Aihua Mao, Zihui Du, Junhui Hou, Yaqi Duan, Yong-jin Liu and Ying He

Introduction

Official PyTorch implementation of our TVCG paper: [Paper & Supplement]

Environment

First clone the code of this repo:

git clone --recursive https://github.com/unknownue/puflow

Then other settings can be either configured manually or set up with docker.

Manual configuration

The code is implemented with CUDA 11.1, Python 3.8, PyTorch 1.8.0.

Other require libraries:

• pytorch-lightning==1.2.2 for training
• tensorflow==1.14.0 for dataset loading
• torchdiffeq form https://github.com/rtqichen/torchdiffeq
• pytorch3d from https://github.com/facebookresearch/pytorch3d
• knn_cuda from https://github.com/unlimblue/KNN_CUDA

Docker configuration

If you are familiar with Docker, you can use provided Dockerfile to configure all setting automatically.

Additional configuration for training

If you want to train the network, you also need to build the kernel of emd like followings:

cd metric/emd/
python setup.py install --user

Datasets

All training and evaluation data can be downloaded from this link, including:

• Training data from PUGeo dataset (tfrecord_x4_normal.zip), PU-GAN dataset and PU1K dataset. Put training data as list in here.
• Testing models of input 2K/5K points and corresponding ground truth 8K/20K points.
• Training and testing meshes for further evaluation.

We include some pretrained x4 models in this repo.

Training & Upsampling & Evaluation

Train the model on specific dataset:

python modules/discrete/train_pu1k.py      # Train the discrete model on PU1K Dataset
python modules/discrete/train_pugeo.py     # Train the discrete model on PUGeo Dataset
python modules/discrete/train_pugan.py     # Train the discrete model on PU-GAN Dataset
python modules/continuous/train_interp.py  # Train the continuous model on PU1K Dataset

Upsampling point clouds as follows:

# For discrete model
python modules/discrete/upsample.py \
    --source=path/to/input/directory \
    --target=path/to/output/directory \
    --checkpoint=pretrain/puflow-x4-pugeo.pt \
    --up_ratio=4

# For continuous model
python modules/continuous/upsample.py \
    --source=path/to/input/directory \
    --target=path/to/output/directory \
    --checkpoint=pretrain/puflow-x4-cnf-pu1k.pt \
    --up_ratio=4

Evaluation as follows:

# Build files for evaluation (see build.sh for more details)
bash evaluation/build.sh

# Evaluate on PU1K dataset
cd evaluation/
cp path/to/output/directory/**.xyz ./result/
bash eval_pu1k.sh
python evaluate.py --pred ./result/ --gt=../data/PU1K/test/input_2048/gt_8192 --save_path=./result/

# Evaluate on PU-GAN dataset
cd evaluation/
cp path/to/output/directory/**.xyz ./result/
bash eval_pugan.sh
python evaluate.py --pred ./result/ --gt=../data/PU-GAN/GT --save_path=./result/

Citation(TODO)

If this work is useful for your research, please consider citing:

@article{unknownue2022puflow,
    title={PU-Flow: a Point Cloud Upsampling Network with Normalizing Flows},
    author={Aihua Mao and Zihui Du and Junhui Hou and Yaqi Duan and Yong-jin Liu and Ying He},
    journal={IEEE Transactions on Visualization and Computer Graphics},
    volume={},
    number={},
    pages={},
    year={2022},
    doi={}
}