single image depth estimation based fourier domain analysiscvpr18基于域分析单幅图像深度估计.pdfVIP

Single-ImageDepthEstimationBasedonFourierDomainAnalysis

Jae-HanLee,MinhyeokHeo,Kyung-RaeKim,andChang-SuKim

KoreaUniversity

{jaehanlee,mhheo,krkim}@mcl.korea.ac.kr,changsukim@korea.ac.kr

incertainperspective[9],orthatascenehasthefloor-walls

geometry[4].Also,focalblurswereexploitedin[41],and

Weproposeadeeplearningalgorithmforsingle-imagehazestrengthsinferredfromthedarkchannelpriorwere

depthestimationbasedontheFourierfrequencydomainusedin[9].Thesetechniques,however,canreconstruct

analysis.First,wedevelopaconvolutionalneuralnetworkdepthsinspecificcasesonlywhenthecorrespondingas-

structureandproposeanewlossfunction,calleddepth-sumptionsarevalid.

balancedEuclideanloss,totrainthenetworkreliablyforRecently,severalmethodshavebeenproposedtoem-

awiderangeofdepths.Then,wegeneratemultipledepthploygraph-basedmodels[21,23,29,30],suchasMarkov

mapcandidatesbycroppinginputimageswithvariousrandomfield(MRF)andconditionalrandomfield(CRF).

croppingratios.Ingeneral,acroppedimagewithasmallMoreover,withthefastdevelopmentofdeeplearningtech-

ratioyieldsdepthdetailsmorefaithfully,whilethatwithanology[11,12,16],variousattemptshavebeenmadeto

largeratioprovidestheoveralldepthdistributionmorere-useconvolutionalneuralnetworks(CNNs)forsingle-image

liably.Totakeadvantageofthesecomplementaryproper-depthestimation