PKSS-Align: A New Registration Method

It is an implementation for the paper “PKSS-Align”. The source code will be released when the paper is accepted. The complied version and related exe file are released at first for testing.

Introduction

This work is an improved version based on KSS-ICP. It can be used to register point clouds with large difference of poses, different scales, noisy points, and defective parts at the same time. With a well designed GPU-based structure, the computational efficiency is improved obviously.

Citation

If you find our work useful for your research, please consider citing:

@article{lv2023kss,
     title={KSS-ICP: Point Cloud Registration Based on Kendall Shape Space},
     author={Lv, Chenlei and Lin, Weisi and Zhao, Baoquan},
     journal={IEEE Transactions on Image Processing},
     volume={32},
     pages={1681--1693},
     year={2023},
     publisher={IEEE}
}

Comparsions

In this part, we introduce some selected methods to estbalish comparsions in different registration tasks. Test datasets and codes of all selected methods are provided in this repository. Some registration instances and related quantitative anlayis reports are shown at the same time. We hope such data can help some ones who are focus on point cloud-based registration research work.

Selected Methods:

Paper	Method	Type	Published	Code	Keywords
pdf	ICP	Distance Metric	TPAMI1992	PCL	iterative closest point
pdf	NDT	Feature-based	IROS2003	PCL	normal distributions transform
pdf	FPFH	Feature-based	ICRA2009	PCL	local descriptor, FEATURE HISTOGRAMS
arvix	Go-ICP	Distance Metric	TPAMI2016	Official	ICP, Global Searching
cvf	PointNetLK	Deep Learning	CVPR2019	Github	PointNet, Lucas&Kanade
arvix	Fast-ICP	Distance Metric	TPAMI2022	Official	Majorization-minimization, Welsch’s function
arvix	Robust-ICP	Distance Metric	TPAMI2022	Official	Majorization-minimization, Welsch’s function
cvf	Geo-Transformer	Deep Learning	CVPR2022	Official	Transformer, RANSAC
cvf	Maximal Cliques	Feature-based	CVPR2023	Official	Maximal Cliques (Best Student Paper)
pdf	KSS-ICP	Distance Metric	TIP2023	Official	Kendall Shape Space

Visualization

Test Report:

We test selected methods and PKSS-Align in the two datasets: ModelNet40 and S3DIS.

The registration test datasets have been prepared (here), which contains different source point clouds with various influnece factors, template point clouds, and related ground truth transformations.

The quantitative anlaysis results are shown in following tables:

Table1: Test Report on ModelNet40 with similarity transformations:

Method	Time	MSE	MSE(n)	GT_cos
ICP	1.981s	0.01253	0.6915	0.4259
NDT	3.719s	0.01157	0.6735	0.4079
FPFH	16.665s	0.00106	0.2163	0.6837
Go-ICP	33.101s	0.00014	0.0885	0.7824
PointNetLK	0.903s	0.02621	0.7476	0.3871
Fast-ICP	6.798s	0.00808	0.5255	0.4668
Robust-ICP	22.921s	0.01207	0.4849	0.4838
Geo-Transformer	0.563s	0.09309	0.8018	0.4776
Maximal-Cliques	2.513s	0.02898	0.4558	0.6688
KSS-ICP	2.899s	0.00033	0.1233	0.7648
PKSS-Align	2.938s	0.00051	0.0589	0.9026