用于非刚性点云匹配神固有嵌入方法研究.pdfVIP

NeuralIntrinsicEmbeddingforNon-rigidPointCloudMatching

PuhuaJiang1,2MingzeSun1RuqiHuang1

1TsinghuaShenzhenInternationalGraduateSchool

2PengChengLaboratory

{jph21,smz22}@mails.tsinghua.edu.cnruqihuang@sz.tsinghua.edu.cn

3

2

0

2

r

a

M

2

]

V

C

.

s

cFigure1.Givenapointcloud,weselect5landmarks(seeredpointsontheright-mostone)andassigneachoftherestpointstothe

[clusterrepresentedbyitsnearestneighboramongthelandmarksintherespectiveembeddedspace.WecompareourmethodtoEuclidean

coordinates,LIE[30],GPS[42].Ourmethodtakesinonlypointcloudandproducessegmentationthatisintrinsicgeometry-aware.

1

v

8

3manycomputervisionandgraphicsapplications,including

0animation[22,36],robotics[16,46],autonomousdriv-

1Asaprimitive3Ddatarepresentation,pointcloudsing[11,55],tonameafew.Incontrasttothewellknown

0.areprevailingin3Dsensing,yetshortofintrinsicstruc-rigidcase,moresophisticateddeformationmodelsarein

3turalinformationoftheunderlyingobjects.Suchdiscrep-demandtocharacterizethenon-rigidmotions,forinstance,

0ancyposesgreatchallengesondirectlyestablishingcorre-articulationmovementsofhumanshapes.

3spondencesweenpointcloudssampledfromdeformableToaddressthischallenge,extrinsicmethodsinprinci-

2

:shapes.Inlightofthis,weproposeNeuralIntrinsicEmbed-pleapproximateacomplexglobalnon-rigiddeformation

v