Learning Graph-Convolutional Representations for Point Cloud Denoising (ECCV 2020)

Bibtex entry:

@inproceedings{pistilli2020learning,
  title={Learning Graph-Convolutional Representationsfor Point Cloud Denoising},
  author={Pistilli, Francesca and Fracastoro, Giulia and Valsesia, Diego and Magli, Enrico},
  booktitle={The European Conference on Computer Vision (ECCV)},
  year={2020}
}

Requirements

• Python 2.7
• Tensorflow 1.12 with CUDA 9.0 (warning: later versions of tensorflow/cuda might have numerical issues)
• point_cloud_utils (https://github.com/fwilliams/point-cloud-utils)
• h5py
• numpy
• scipy

Code structure

Code/ : Python source code for training/testing and the network model Dataset/ : Shapenet training and testing data log_dir/ : log files and tensorboard events Results/ : saved checkpoints and denoised point clouds

Dataset

Download the Dataset directory from: https://www.dropbox.com/sh/nwdlzgnt987yjma/AAAi8q0E6yioxk5I_BkUpZE5a?dl=0

Pretrained models

Pretrained models are included for the MSE-SP loss and 16-NN at standard deviations 0.01,0.015,0.02. Results might be slightly different from the ones in the paper.

Test

./launcher_test.sh

Denoised point clouds and C2C metrics will be written to Results directory.

Train

./launcher_train.sh

Checkpoints will be written to the Results directory.

Notes

Code is partially based on this project (https://github.com/diegovalsesia/gcdn). Check out its documetation for more details on the parameters in config.py. Training and tested was run on

CPU: AMD Ryzen 1 1700
RAM: 32 GB
GPUs: 1x Nvidia Quadro P6000 (24 GB)