人工智能论文20230311 2022PhD Virginia Tech_3D Deep Learning for Object-Centric Geometric Perception.pdfVIP

3DDeepLearningforObject-CentricGeometricPerception

XiaolongLi

DissertationsubmittedtotheFacultyofthe

VirginiaPolytechnicInstituteandStateUniversity

inpartialfulfillmentoftherequirementsforthedegreeof

DoctorofPhilosophy

in

ComputerEngineering

A.LynnAbbott,Chair

Jia-BinHuang

RyanWilliams

NicholasF.Polys

WeiZhou

May6th,2022

Blacksburg,Virginia

Keywords:pointcloud,poseestimation,equivariance,motionforecasting,shapecompletion

Copyright2022,XiaolongLi

3DDeepLearningforObject-CentricGeometric

Perception

XiaolongLi

ABSTRACT

Object-centricgeometricperceptionaimsatextractingthegeometricattributesof3Dob-

jects.Theseattributesincludeshape,pose,andmotionofthetargetobjects,whichenable

fine-grainedobject-levelunderstandingforvarioustasksingraphics,computervision,and

robotics.Withthegrowthof3Dgeometrydataand3Ddeeplearningmethods,itbecomes

moreandmorelikelytoachievesuchtasksdirectlyusing3Dinputdata.Amongdifferent3D

representations,a3Dpointcloudisasimple,common,andmemory-efficientrepresentation

thatcouldbedirectlyretrievedfrommulti-viewimages,depthscans,orLiDARrangeimages.

Differentchallengesexistinachievingobject-centricgeometricperception,suchasachieving