Stereo-Matching-by-Training-a-Convolutional-Neural-Network-to-Compare-Image-Patches

《Stereo-Matching-by-Training-a-Convolutional-Neural-Network-to-Compare-Image-Patches》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、StereoMatchingbyTrainingaConvolutionalNeuralNetworktoCompareImagePatchesJureZbontarˇZBONTAR@CS.NYU.EDUYannLeCunyYANN@CS.NYU.EDUCourantInstituteofMathematicalSciencesNewYorkUniversity715Broadway,NewYork,NY10003,USAAbstractWepresentamethodforextractingdepthinformationf

2、romarectifiedimagepair.Ourapproachfocusesonthefirststageofmanystereoalgorithms:thematchingcostcomputation.Weapproachtheproblembylearningasimilaritymeasureonsmallimagepatchesusingaconvolutionalneuralnetwork.Trainingiscarriedoutinasupervisedmannerbyconstructingabinaryclas

3、sificationdatasetwithexamplesofsimilaranddissimilarpairsofpatches.Weexaminetwonetworkarchitec-turesforthistask:onetunedforspeed,theotherforaccuracy.Theoutputoftheconvolutionalneuralnetworkisusedtoinitializethestereomatchingcost.Aseriesofpost-processingstepsfol-low:cros

4、s-basedcostaggregation,semiglobalmatching,aleft-rightconsistencycheck,subpixelenhancement,amedianfilter,andabilateralfilter.WeevaluateourmethodontheKITTI2012,KITTI2015,andMiddleburystereodatasetsandshowthatitoutperformsotherapproachesonallthreedatasets.Keywords:stereo,m

5、atchingcost,similaritylearning,supervisedlearning,convolutionalneuralnetworks1.IntroductionConsiderthefollowingproblem:giventwoimagestakenfromcamerasatdifferenthorizontalpo-sitions,wewishtocomputethedisparitydforeachpixelintheleftimage.Disparityreferstothedifferencein

6、horizontallocationofanobjectintheleftandrightimage—anobjectatposition(x;y)intheleftimageappearsatposition(xd;y)intherightimage.Ifweknowthedisparityofanobjectwecancomputeitsdepthzusingthefollowingrelation:arXiv:1510.05970v1[cs.CV]20Oct2015fBz=;dwherefisthefocallengtho

7、fthecameraandBisthedistancebetweenthecameracenters.Figure1depictstheinputtoandtheoutputfromourmethod.Thedescribedproblemofstereomatchingisimportantinmanyfieldssuchasautonomousdriving,robotics,intermediateviewgeneration,and3Dscenereconstruction.Accordingtothetax-onomyof

8、ScharsteinandSzeliski(2002),atypicalstereoalgorithmconsistsoffoursteps:matching.JureZbontarisalsowiththeˇFacultyofComputera