Self-Supervised Arbitrary-Scale Point Clouds Upsampling via Implicit Neural Representation (SAPCU)
This is the code project of Self-Supervised Arbitrary-Scale Point Clouds Upsampling via Implicit Neural
Representation published on CVPR 2022.
The origin project is based on pytorch and we provide the tensorflow version in ./torch.
Environment
Python 3.6
PyTorch 1.9.0
Tensorflow 2.3
Tensorlayer3
CUDA 10.2
Note That
you can run the origin pytorch project by the follow commands, while you should
cd ./torch
, to train or test the transplanted tensorflow project.
Evaluation
The pretrained models are put in dir ./out.
Firstly, you should compile cpp
g++ -std=c++11 dense.cpp -O2 -o dense
Then, you can run
python generate.py
Trainning
Download the training dataset from the link and unzip it to the working directory.
https://pan.baidu.com/s/1VQ-3RFO02fQfcLBfqvCBZA
access code: vpfm
Then run the following commands for training sapcu network
python trainfn.py
python trainfd.py
Performance
The 4X upsampling performance on pytorch and tensorflow platform is statisted in follow table,
PyTorch
CD |
EMD |
F-score |
AVG |
STD |
0.010127 |
0.004974 |
0.561682 |
0.003492 |
0.009351 |
Tensorflow
CD |
EMD |
F-score |
AVG |
STD |
0.012917 |
0.009506 |
0.472918 |
0.008855 |
0.018178 |
Cite
Wenbo Zhao, Xianming Liu, Zhiwei Zhong, Junjun Jian, Wei Gao, Ge Li, Xiangyang Ji, “Self-Supervised Arbitrary-Scale Point Clouds Upsampling via Implicit Neural Representation,” Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR) 2022.
contributors
name: Zhang Yongchi
email: zhangych02@pcl.ac.cn