资源描述:
《Efficient Sequential Correspondence Selection by Cosegmentation》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、1568IEEETRANSACTIONSONPATTERNANALYSISANDMACHINEINTELLIGENCE,VOL.32,NO.9,SEPTEMBER2010EfficientSequentialCorrespondenceSelectionbyCosegmentationJanCech,Member,IEEE,Jirı´Matas,Member,IEEE,andMichalPerdoch,Member,IEEEAbstract—Inmanyretrieval,objectrecognition,andw
2、ide-baselinestereomethods,correspondencesofinterestpoints(distinguishedregions)arecommonlyestablishedbymatchingcompactdescriptorssuchasSIFTs.Weshowthatasubsequentcosegmentationprocesscoupledwithaquasi-optimalsequentialdecisionprocessleadstoacorrespondenceverifica
3、tionprocedurethat1)hashighprecision(ishighlydiscriminative),2)hasgoodrecall,and3)isfast.Thesequentialdecisiononthecorrectnessofacorrespondenceisbasedonsimplestatisticsofamodifieddensestereomatchingalgorithm.Thestatisticsareprojectedonaprominentdiscriminativedirec
4、tionbySVM.Wald’ssequentialprobabilityratiotestisperformedontheSVMprojectioncomputedonprogressivelylargercosegmentedregions.Weshowexperimentallythattheproposedsequentialcorrespondenceverification(SCV)algorithmsignificantlyoutperformsthestandardcorrespondenceselect
5、ionmethodbasedonSIFTdistanceratiosonchallengingmatchingproblems.IndexTerms—Correspondence,matching,verification,sequentialdecision,growing,cosegmentation,stereo,imageretrieval,learning.Ç1INTRODUCTIONANYsuccessfulimageretrieval,objectrecognition,andgeometricnormal
6、ization.Additionally,thedescriptormayMwide-baselinestereomethodsexploitcorrespon-becompressedbyquantization.1dencesofdistinguishedregions.Mostreal-worldvisualThisprocess,schematicallyvisualizedinFig.2,hastherecognitionproblemsarelargescale,wherecorrespon-followin
7、gmaincharacteristics:1)Allstepsareperformedindencesbetweenregionsfromaquery(test)imageandmanyindividualimagesindependently,2)theshapeandsizeofdatabase(training)imagesofobjectsorscenesaresought.themeasurementregionareafixedfunctionoftheshapeToachieveacceptableresp
8、onsetimes,largeproblemsandsizeofthedistinguishedregion,and3)thedescriptorrequirethetimecomplexityoftheregionmatchingprocesshasthesameformforallregions,e.g.,iti