<?xml version="1.0" encoding="ISO-8859-1"?>
<metadatalist>
	<metadata ReferenceType="Conference Proceedings">
		<site>sibgrapi.sid.inpe.br 802</site>
		<identifier>8JMKD3MGPEW34M/49SP3UH</identifier>
		<repository>sid.inpe.br/sibgrapi/2023/09.26.00.25</repository>
		<lastupdate>2024:01.31.12.51.34 sid.inpe.br/banon/2001/03.30.15.38 gmnetto@inf.ufrgs.br</lastupdate>
		<metadatarepository>sid.inpe.br/sibgrapi/2023/09.26.00.25.49</metadatarepository>
		<metadatalastupdate>2024:01.31.12.51.34 sid.inpe.br/banon/2001/03.30.15.38 gmnetto@inf.ufrgs.br {D 2023}</metadatalastupdate>
		<citationkey>Netto:2023:RoPoRe</citationkey>
		<title>Robust Point-Cloud Registration based on Dense Point Matching and Probabilistic Modeling</title>
		<format>On-line</format>
		<year>2023</year>
		<numberoffiles>1</numberoffiles>
		<size>3816 KiB</size>
		<author>Netto, Gustavo Marques,</author>
		<affiliation>UFRGS</affiliation>
		<editor>Clua, Esteban Walter Gonzalez,</editor>
		<editor>Körting, Thales Sehn,</editor>
		<editor>Paulovich, Fernando Vieira,</editor>
		<editor>Feris, Rogerio,</editor>
		<e-mailaddress>gmnetto@inf.ufrgs.br</e-mailaddress>
		<conferencename>Conference on Graphics, Patterns and Images, 36 (SIBGRAPI)</conferencename>
		<conferencelocation>Rio Grande, RS</conferencelocation>
		<date>Nov. 06-09, 2023</date>
		<booktitle>Proceedings</booktitle>
		<tertiarytype>Master's or Doctoral Work</tertiarytype>
		<transferableflag>1</transferableflag>
		<keywords>Point-cloud registration, rigid registration, non-rigid registration ,dense point matching.</keywords>
		<abstract>This thesis presents techniques for 3D point-cloud registration that are robust to outliers and missing regions. They tackle non-rigid and rigid registration and exploit the advantages of deep learning for dense point matching. This is done by proposing a single new neural network to solve both registration types. Our network uses a recently proposed attention mechanism and explicitly accounts for missing correspondences, which is key to its performance. Additionally, we use recent advances in probabilistic modeling to further refine the correspondences created by our network during non-rigid registration. Such a combination of deep learning and probabilistic modeling produces context awareness and enforces motion coherence, which makes our approach resilient to outliers and missing information. We demonstrate the effectiveness of our techniques by comparing them to state-of-the-art methods. Our comparisons use datasets containing noise, partial point clouds, and irregular sampling. The experiments show that our techniques obtain superior results in general. For example, our approaches achieve a registration error up to 45% smaller than other techniques in partial point clouds for non-rigid registration, and up to 49% smaller on rigid registration. We also discuss additional aspects of our techniques such as robustness to different levels of noise and to different numbers of samples in the point clouds. Finally, we tackle the lack of datasets with ground truth for supervised training of non-rigid registration models by presenting a self-supervised strategy based on random deformations.</abstract>
		<language>en</language>
		<targetfile>SIBGRAPI2023_Netto-1-1.pdf</targetfile>
		<usergroup>gmnetto@inf.ufrgs.br</usergroup>
		<visibility>shown</visibility>
		<mirrorrepository>sid.inpe.br/banon/2001/03.30.15.38.24</mirrorrepository>
		<hostcollection>sid.inpe.br/banon/2001/03.30.15.38</hostcollection>
		<username>gmnetto@inf.ufrgs.br</username>
		<agreement>agreement.html .htaccess .htaccess2</agreement>
		<lasthostcollection>sid.inpe.br/banon/2001/03.30.15.38</lasthostcollection>
		<url>http://sibgrapi.sid.inpe.br/rep-/sid.inpe.br/sibgrapi/2023/09.26.00.25</url>
	</metadata>
</metadatalist>