人工智能核心550篇论文 3D重建 1908.07906v2.pdfVIP

PCRNet:PointCloudRegistrationNetworkusingPointNetEncoding

VinitSarode1XueqianLiHunterGoforth3YasuhiroAoki2

RangaprasadArunSrivatsan4SimonLucey1,3HowieChoset1

1CarnegieMellonUniversity2FujitsuLaboratoriesLtd.3ArgoAI.4Apple.

vsarode,xueqianl@aoki-yasuhiro@

9hgoforth,slucey,choset@aruns@

1

0

2

vAbstract

o

NPointNethasrecentlyemergedasapopularrepresenta-

4tionforunstructuredpointclouddata,allowingapplication

ofdeeplearningtotaskssuchasobjectdetection,segmen-

]tationandshapecompletion.However,recentworksinlit-

VeraturehaveshownthesensitivityofthePointNetrepresen-

Ctationtoposemisalignment.Thispaperpresentsanovel

s.frameworkthatusesthePointNetrepresentationtoalign

cpointcloudsandperformregistrationforapplicationssuch

[

astracking,3Dreconstructionandposeestimation.Wede-

2velopaframeworkthatcomparesPointNetfeaturesoftem-Figure1:Comparisonofdifferentregistrationmethods

vplateandsourcepointcloudstofindthetransformationthatbasedontheirrobustnesstonoiseandcomputationtime

6alignsthemaccurately.Dependingonthepriorinformationwithrespecttoobjectspecificity.Theiterativeversionof

0

9abouttheshapeoftheobjectformedbythepointclouds,PCRNetexploitsobjectspecificitytoproduceaccuratere-

7ourframeworkcanproduceapproachesthatareshapespe-sults.The