RePCD-Net: Feature-aware Recurrent Point Cloud Denoising Network

Introduction

This repository is for our International Journal of Computer Vision (IJCV) 2022 paper 'RePCD-Net: Feature-aware Recurrent Point Cloud Denoising Network'.

Denoised result

We have released some denoised results in our work, please feel free to use them.

Synthetic test dataset

We have aslo released our synthetic test dataset for a easiser comparison for future researchers. For the quantitative statistics, please refer to the table 2 in this paper. Note also that this dataset is built based on the 'PU-GAN'.

Taining dataset

Download the training dataset train_4000_normal_scale_label_weight_61_6.h5 from here. Then put it in the folder ../h5_data.

Usage

1. Clone the repository:

   git clone https://github.com/chenhonghua/Re-PCD.git
   cd Re-PCD

2. Compile the TF operators Follow the above information to compile the TF operators.

3. train the model: run:

   cd codes
   python main.py --phase train

4. Evaluate the model: run:

   cd codes
   python main.py --phase test

   You will see the input and output results in the folder ../data/test_data and ../model/generator2_new6/result/.

Note: During the test stage, we consider the entire input point cloud as a single entity. However, if the input point cloud contains a large number of points, it is advisable to partition it into smaller patches and process each patch individually as separate inputs.

Citation

If you use this dataset, please consider citing our work.

@article{chen2022repcd,
title={RePCD-Net: Feature-Aware Recurrent Point Cloud Denoising Network},
author={Chen, Honghua and Wei, Zeyong and Li, Xianzhi and Xu, Yabin and Wei, Mingqiang and Wang, Jun},
journal={International Journal of Computer Vision},
pages={1--15},
year={2022},
publisher={Springer}
}