xucao-42/psnet

Official implementation of "PSNet: A Style Transfer Network for Point Cloud Stylization on Geometry and Color, " WACV 2020.

Code for WACV2020 paper "PSNet: A Style Transfer Network for Point Cloud Stylization on Geometry and Color"

Introduction

We perform neural style transer on a point cloud from a point cloud or an image from a style point cloud. The geometry or/and the color property of the content point cloud can be stylized. The color propety can also be stylized from an image.

Usage

Dependencies

Code was tested on MacOS 10.14 & above and Ubuntu 16.04 with Python3.6.

Install required packages: pip3 install -r requirements.txt.

Be cautious not upgrading matplotlib to 3.1.0 or above, it will drop an error message "It is not currently possible to manually set the aspect on 3D axes" when visualizing point clouds.

Basic usage

Just run main.py. It will style transfer all point clouds in sample_content from each style image or point cloud in sample_style. Results are saved in a new folder style_transfer_results.

Prepare your data

Put your .ply content point clouds in sample_content and your style images or point clouds in sample_style. Then run main.py.

Citation

If you used this code in your publication, please consider citing the following paper:

@InProceedings{Cao_2020_WACV,
author = {Cao, Xu and Wang, Weimin and Nagao, Katashi and Nakamura, Ryosuke},
title = {PSNet: A Style Transfer Network for Point Cloud Stylization on Geometry and Color},
booktitle = {The IEEE Winter Conference on Applications of Computer Vision (WACV)},
month = {March},
year = {2020}
}

Contact

For any questions/comments/bug reports, please feel free to contact [email protected]

psnet's Issues

Training code & train using my own data

Hi,

I am very interested in your work. From the code available (i.e. main.py), I found that it loads weights from a trained network model and directly tests on the input point clouds. However, I wish to retrain your network on my own dataset. Is the code for training available in this GitHub repository? If not, will you release it? I am really curious about how it'll look like after being trained on my own dataset.

Regards,
Wei