Retrieval-and-alignment based large-scale indoor point cloud semantic segmentation  

作　　者：Zongyi XU Xiaoshui HUANG Bo YUAN Yangfu WANG Qianni ZHANG Weisheng LI Xinbo GAO 

机构地区：[1]School of Computer Science and Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,China [2]Chongqing Institute for Brain and Intelligence,Guangyang Bay Laboratory,Chongqing 400064,China [3]Shanghai Artificial Intelligence Laboratory,Shanghai 200232,China [4]School of Electronic Engineering and Computer Science,Queen Mary University of London,London E14NS,UK

出　　处：《Science China(Information Sciences)》2024年第4期160-176,共17页中国科学（信息科学）（英文版）

基　　金：supported by National Natural Science Foundation of China(Grant Nos.62206033,62221005,U22A2096);Natural Science Foundation of Chongqing(Grant Nos.cstc2020jcyj-msxmX0855,cstc2021ycjh-bgzxm0339);Chongqing Postdoctoral Research Special Funding Project(Grant No.2021XM2044)。

摘　　要：Current methods for point cloud semantic segmentation depend on the extraction of descriptive features.However,unlike images,point clouds are irregular and often lack texture information,making it demanding to extract discriminative features.In addition,noise,outliers,and uneven point distribution are commonly present in point clouds,which further complicates the segmentation task.To address these problems,a novel architecture is proposed for direct and accurate large-scale point cloud segmentation based on point cloud retrieval and alignment.The proposed approach involves using a feature-based point cloud retrieval method for searching for reference point clouds with annotations from a dataset.In the following segmentation stage,an overlap-based point cloud registration method has been developed to align the target and reference point clouds.For accurate and robust alignment,an overlap region estimation module is trained to locate the optimal overlap region between two pieces of point clouds in a coarse-to-fine manner.In the detected overlap region,the global and local features of the points are extracted and combined for featuremetric registration to obtain accurate transformation parameters between the target and reference point clouds.After alignment,the annotated segmentation of the reference is transferred to the target point clouds to obtain accurate segmentation results.Extensive experiments are conducted to show that the developed method outperforms the state-of-the-art approaches in terms of both accuracy and robustness against noise and outliers.

关 键 词：point cloud semantic segmentation large-scale indoor point clouds point cloud alignment overlap estimation label transfer 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]