資源描述：
《Domain Adaptive Object Detection》由會(huì)員上傳分享，免費(fèi)在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫(kù)。
1、DomainAdaptiveObjectDetectionFatemehMirrashed1,VladI.Morariu1,BehjatSiddiquie2,RogerioS.Feris3,LarryS.Davis11UniversityofMaryland,CollegePark2SRIInternational3IBMResearchffatemeh,morariu,lsdg@umiacs.umd.edubehjat.siddiquie@sri.comrsferis@us.ibm.comAbstractWestudythe
2、useofdomainadaptationandtransferlearningtechniquesaspartofaframeworkforadaptiveob-jectdetection.Unlikerecentapplicationsofdomainadap-tationworkincomputervision,whichgenerallyfocusonimageclassi?cation,weexploretheproblemofextremeclassimbalancepresentwhenperformingdom
3、ainadapta-tionforobjectdetection.Themaindif?cultycausedbythisimbalanceisthattestimagescontainmillionsorbillionsofnegativeimagesubwindowsbutjustafewpositiveones,whichmakesitdif?culttoadapttothechangesinthepos-itiveclassdistributionsbysimpletechniquessuchasran-Figure1
4、.Anexampleoftheeffectsofdomainchangeforthetaskofvehicledetectionandourimprovedresultsafterdomainadap-domsampling.Weproposeaninitialapproachtoaddresstation.Here,thevehicledetectoristrainedontrainingdata,thethisproblemandapplyourtechniquetovehicledetectionsourcedomain
5、,andisappliedtotestingdata(anewdomain)thatinachallengingurbansurveillancedataset,demonstratingdiffersfromthetrainingdatainvariousways,e.g.,viewingangles,theperformanceofourapproachwithvariousamountsofillumination.Ifwedirectlyapplythetrainedmodeltoanewdo-supervision,
6、includingthefullyunsupervisedcase.main,thecon?dencemaphasmultiplepeaks,manyofwhichdonotcorrespondtovehicles.Afterdomainadaptation,thehighestpeakscorrespondtothetwovehiclesintheforeground.(Note:1.IntroductionBackgroundregionshavebeenobfuscatedforlegal/privacyrea-sons
7、)Buildingvisualmodelsofobjectsrobusttoextrinsic1variationssuchascameraviewangle(orobjectpose),reso-lution,lighting,andblurhaslongbeenoneofthechallengesmechanismstotransferoradaptknowledgefromonedo-incomputervision.Generally,adiscriminativeorgenera-maintoanotherrelat
8、eddomain.Whiletheseadvanceshavetivestatisticalmodelistrainedbyacquiringalargesetofex-alsobeenappliedbythecomputervisioncommunitywithamples
