基于ℓ1稀疏重建尖锐点集表面处理技术.pptx

?1-sparseReconstruction

ofSharpPointSetSurfacesHaimAvron,AndreiSharf,ChenGreif,DanielCohen-Or

MotivationPiecewisesmoothshape/fewsharpregionsSparsesetoffeaturesSparsesignal

reconstruction

GivenadiscretesignalSignalreconstructionSparsestsolution(Sparse)SignalReconstruction1

?1normissparse

OurIdeaReconstructnoisypointsutilizing?1:avoidslocalminima/yieldsasparsesolutionavoidssmoothederror(L2)

SparseGradientsin2D(TV)PhotoGradientCleanNoisy

TotalVariation(2D)[RUDIN,OSHER,FATEMI92]Localgradients:Constrained?1minimization:

?1SparseReconstructionnormalspositionsNoisyinputSparseorientationsPositions2DV-shapeexample

signal?2?1input140’150’160’?1vs.?2

?1concentrates/?2propagateserrors

?1SparseReconstructionofFanDisksmoothsharpShapeReconstructedNoisyNormalvariation

?1SparseReconstructionofScannedBudhaShapeReconstructedsmoothsharpInputNormalvariation

RelatedWorkSparsesignalreconstruction

[Totalvariation92][Basispursuit01][Totalcurvature2009]Featureawarelocalfilters

[Bilateraldenoising03][RobustMLS05][DDMLS07][LOP07][RIMLS09]below?1dynamicscenereconstruction

[4Dpressibleflow08]Lipmanet.al.07Tasdizenet.al.03Sharfet.al.08

sharppointnormaldifference?1-minimizationEstimateshapesmoothnessusingnormaldifferencesGlobaloptimizationproblemsmooth

positionprojection?1-minimizationSmoothregionsareapproximatedwellbynormalsProjectiondistanceofpointstolocalplanesmoothsharpfeature

Thegood,thebadandtheuglyProblemisconvex:aglobalminimum?1normnotdifferentiableduetoabsolutevaluesAnontrivialiterativesolverisrequired

RecastingCannotworkwithabsolutesAddvariableu,andminimizeRequireTrueif

RecastLinearobjectiveConvexconstraintsSecondorderconeproblem(SOCP)

VarioussolversProjectedgradientsSPGL1–vandenBergandFriedlander,GPSR–Figueiredo,NowakandWright…Interior-pointmethodsl1-magic–Cand