PU-EdgeFormer: Edge Transformer for Dense Prediction in Point Cloud Upsampling

by Dohoon Kim, Minwoo Shin, and Joonki Paik.

This is the official implementation of PU-Edgeformer: Edge Transformer for Dense Prediction for Point Cloud Upsampling.

This repository supports training our paper PU-EdgeFormer, and previous methods PU-Net, MPU, PU-GAN, PU-GCN.

Installation

git clone https://github.com/dohoon2045/PU-EdgeFormer.git
cd puedgeformer
bach env_install.sh
conda activate puedgeformer

Dataset

We use PU1K dataset for training and testing as provided by PU-GCN. Please refer to original repository for downloading the data.

You can also use other dataset of h5 format such as provided by PU-GAN, PU-Net.

Training

python main.py --phase train --model puedgeformer --log_dir log/pu-edgeformer/

Testing

python main.py --phase test --model puedgeformer --log_dir log/pu-edgeformer/ --data_dir ./data/PU1K/test/input_2048/input_2048/

Evaluation

python evaluate.py --gt ./data/PU1K/test/input_2048/gt_8192/ --pred evaluation_code/result/ --save_path log/pu-edgeformer/

Citation

If you find PU-EdgeFormer is useful your research, please consider citing:

@article{kim2023pu,
title={PU-EdgeFormer: Edge Transformer for Dense Prediction in Point Cloud Upsampling},
author={Kim, Dohoon and Shin, Minwoo and Paik, Joonki},
journal={arXiv preprint arXiv:2305.01148},
year={2023}
}

Acknowledgement

This repo is heavily built based on PU-GCN and PU-GAN code. We also borrow the architecture and evaluation codes from PU-Net and MPU.