cvpr18-Unsupervised Person Image Synthesis in Arbitrary Poses

《cvpr18-Unsupervised Person Image Synthesis in Arbitrary Poses》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、UnsupervisedPersonImageSynthesisinArbitraryPosesAlbertPumarolaAntonioAgudoAlbertoSanfeliuFrancescMoreno-NoguerInstitutdeRoboticaiInform`aticaIndustrial(CSIC-UPC)`08028,Barcelona,SpainFigure1:Givenanoriginalimageofaperson(left)andadesiredbodyposedefinedbya2Dskeleton(bottom-row),ourmodelgeneratesn

2、ewphoto-realisticimagesofthepersonunderthatpose(top-row).Themaincontributionofourworkistotrainthisgenerativemodelwithunlabeleddata.Abstract1.IntroductionWepresentanovelapproachforsynthesizingphoto-Beingabletogeneratenovelphoto-realisticviewsofarealisticimagesofpeopleinarbitraryposesusinggener-p

3、ersoninanarbitraryposefromasingleimagewouldopenativeadversariallearning.Givenaninputimageofaper-thedoortomanynewexcitingapplicationsindifferentar-sonandadesiredposerepresentedbya2Dskeleton,oureas,includingfashionande-commercebusiness,photogra-modelrenderstheimageofthesamepersonunderthenewphytec

4、hnologiestoautomaticallyeditandanimatestillim-pose,synthesizingnovelviewsofthepartsvisibleinthein-ages,andthemovieindustrytonameafew.Addressingputimageandhallucinatingthosethatarenotseen.Thisthistaskwithoutexplicitlycapturingtheunderlyingpro-problemhasrecentlybeenaddressedinasupervisedman-cesse

5、sinvolvedintheimageformationsuchasestimatingner[16,35],i.e.,duringtrainingthegroundtruthimagesthe3Dgeometryofthebody,hairandclothes,andtheap-underthenewposesaregiventothenetwork.Wegobe-pearanceandreflectancemodelsofthevisibleandoccludedyondtheseapproachesbyproposingafullyunsupervisedpartsseemsan

6、extremelycomplexendeavor.Nevertheless,strategy.WetacklethischallengingscenariobysplittingGenerativeAdversarialNetworks(GANs)[3]haveshowntheproblemintotwoprincipalsubtasks.First,weconsiderimpressiveresultsinrenderingnewrealisticimages,e.g.,aposeconditionedbidirectionalgeneratorthatmapsbackfaces[

7、8,22],indoorscenes[32]andclothes[39],bydi-theinitiallyrenderedimagetotheoriginalpose,hencebe-rectlylearningagenerativemodelfromdata.Veryrecently,ingdirectlycomparabletotheinputimagewithouttheneedtheyhavebeenusedfortheparticularpro