learning to generate chairs with convolutional neural networks

《learning to generate chairs with convolutional neural networks》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、LearningtoGenerateChairswithConvolutionalNeuralNetworksAlexeyDosovitskiyJostTobiasSpringenbergThomasBroxDepartmentofComputerScience,UniversityofFreiburgfdosovits,springj,broxg@cs.uni-freiburg.deAbstractWetrainagenerativeconvolutionalneuralnetworkwhichisabletogenerateimagesofobject

2、sgivenobjecttype,viewpoint,andcolor.Wetrainthenetworkinasu-pervisedmanneronadatasetofrendered3Dchairmod-els.Ourexperimentsshowthatthenetworkdoesnotmerelylearnallimagesbyheart,butratherfindsameaningfulrepresentationofa3Dchairmodelallowingittoassessthesimilarityofdifferentchairs,inte

3、rpolatebetweengivenviewpointstogeneratethemissingones,orinventnewchairstylesbyinterpolatingbetweenchairsfromthetrainingset.Figure1.Interpolationbetweentwochairmodels(original:topWeshowthatthenetworkcanbeusedtofindcorrespon-left,final:bottomleft).Thegenerativeconvolutionalneuralnet-w

4、orklearnsthemanifoldofchairs,allowingittointerpolatebe-dencesbetweendifferentchairsfromthedataset,outper-tweenchairstyles,producingrealisticintermediatestyles.formingexistingapproachesonthistask.canperfectlyapproximateanyfunctiononthetrainingset.1.IntroductionInourcase,anetworkpot

5、entiallycouldjustlearnbyheartallexamplesandprovideperfectreconstructionsofthese,Convolutionalneuralnetworks(CNNs)havebeenshownbutwouldbehaveunpredictablywhenconfrontedwithin-tobeverysuccessfulonavarietyofcomputervisiontasks,putsithasnotseenduringtraining.Weshowthatthisisnotsuchasi

6、mageclassification[17,5,31],detection[9,27]whatishappening,bothbecausethenetworkistoosmalltoandsegmentation[9].Allthesetaskshaveincommonjustrememberallimages,andbecauseweobservegener-thattheycanbeposedasdiscriminativesupervisedlearn-alizationtopreviouslyunseendata.Namely,weshowthat

7、ingproblems,andhencecanbesolvedusingCNNswhichthenetworkiscapableof:1)knowledgetransfer:givenlim-areknowntoperformwellgivenalargeenoughlabeleditednumberofviewpointsofanobject,thenetworkcanusedataset.Typically,atasksolvedbysupervisedCNNsin-theknowledgelearnedfromothersimilarobjectst

8、oinfervolveslearningmappingsfromr