资源描述:
《苹果人工智能报告:发力图像识别》由会员上传分享,免费在线阅读,更多相关内容在行业资料-天天文库。
1、ThispaperhasbeensubmittedforpublicationonNovember15,2016.LearningfromSimulatedandUnsupervisedImagesthroughAdversarialTrainingAshishShrivastava,TomasPfister,OncelTuzel,JoshSusskind,WendaWang,RussWebbAppleInc.{a_shrivastava,tpf,otuzel,jsusskind,wenda_wang,rwebb}@apple.comAbstractUnl
2、abeledRealImagesWithrecentprogressingraphics,ithasbecomemoretractabletotrainmodelsonsyntheticimages,poten-tiallyavoidingtheneedforexpensiveannotations.How-ever,learningfromsyntheticimagesmaynotachievethedesiredperformanceduetoagapbetweensyntheticandrealimagedistributions.Toreduce
3、thisgap,wepro-poseSimulated+Unsupervised(S+U)learning,whereRefinerthetaskistolearnamodeltoimprovetherealismofasimulator’soutputusingunlabeledrealdata,whileSyntheticRefinedpreservingtheannotationinformationfromthesimula-Figure1.Simulated+Unsupervised(S+U)learning.Thetaskistor.Wedeve
4、lopamethodforS+UlearningthatusestolearnamodelthatimprovestherealismofsyntheticimagesanadversarialnetworksimilartoGenerativeAdversar-fromasimulatorusingunlabeledrealdata,whilepreservingialNetworks(GANs),butwithsyntheticimagesasin-theannotationinformation.putsinsteadofrandomvectors
5、.WemakeseveralkeymodificationstothestandardGANalgorithmtopre-serveannotations,avoidartifactsandstabilizetraining:However,learningfromsyntheticimagescanbeprob-(i)a‘self-regularization’term,(ii)alocaladversariallematicduetoagapbetweensyntheticandrealim-loss,and(iii)updatingthediscri
6、minatorusingahistoryagedistributions–syntheticdataisoftennotrealisticofrefinedimages.Weshowthatthisenablesgenera-enough,leadingthenetworktolearndetailsonlypresenttionofhighlyrealisticimages,whichwedemonstrateinsyntheticimagesandfailtogeneralizewellonrealbothqualitativelyandwithaus
7、erstudy.Wequantita-images.Onesolutiontoclosingthisgapistoimprovetivelyevaluatethegeneratedimagesbytrainingmod-thesimulator.However,increasingtherealismisoftenelsforgazeestimationandhandposeestimation.Wecomputationallyexpensive,therendererdesigntakesashowasignificantimprovementover
8、usingsyntheticim-lotofhardwork,andevento