PUGeo-Net: A Geometry-centric Network for 3D Point Cloud Upsampling

This is the official Tensorflow implementation for paper: https://arxiv.org/abs/2002.10277

The PyTorch version can be found in: https://github.com/rsy6318/PUGeoNet_pytorch

Environment setting

The code is implemented with CUDA=10.0, tensorflow=1.14, python=2.7. Other settings should also be ok.

Other requested libraries: tqdm

Compile tf_ops

One should change the CUDA path in tf_ops/CD/compile.sh and tf_ops/sampling/compile.sh. Then perform

cd tf_ops/CD
sh compile.sh

cd  tf_ops/sampling
sh compile.sh

Some common methods to fix bugs during compiling:

• Make sure you change the CUDA path in compile.sh correctly.
• Make sure you are using (and also compiling under) tensorflow-gpu, not the cpu version of TF.
• You may compile with other TF version. May need to modify the compile.sh. One can refer to the issues of pointnet2, PU-Net, MPU and PU-GAN.
• Delete previous .cu.o and so.so files and recompile again
• Check the "libtensorflow_framework.so" in your tensorflow folder, if it is installed as "libtensorflow_framework.so.1", run this command:

ln -s  libtensorflow_framework.so.1  libtensorflow_framework.so

Datasets and pretrained model

We provide x4 training dataset and pretrained model. Please download these files in the following link:

• training data (tfrecord_x4_normal.zip)
• 13 testing models with 5000 points (test_5000.zip)
• pretrained x4 model (PUGeo_x4.zip)
• Training and testing meshes

https://drive.google.com/drive/folders/1n2lf4am9k3hy3ci4W20XiMkXwJKwyg8f?usp=sharing

Training

python main.py --phase train --up_ratio 4 --log_dir PUGeo_x4

Inference (upsampling)

python main.py --phase test --up_ratio 4 --pretrained PUGeo_x4/model/model-final --eval_xyz test_5000

The upsampled xyz will be stored in PUGeo_x4/eval.

We thank the authors of pointnet2 PU-Net and MPU for their public code.