cvpr18-Wing Loss for Robust Facial Landmark Localisation With Convolutional Neural Networks

《cvpr18-Wing Loss for Robust Facial Landmark Localisation 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、WingLossforRobustFacialLandmarkLocalisationwithConvolutionalNeuralNetworksZhen-HuaFeng1JosefKittler1MuhammadAwais1PatrikHuber1Xiao-JunWu21CentreforVision,SpeechandSignalProcessing,UniversityofSurrey,GuildfordGU27XH,UK2SchoolofIoTEngineering,JiangnanUni

2、versity,Wuxi214122,China{z.feng,j.kittler,m.a.rana}@surrey.ac.uk,patrikhuber@gmail.com,wuxiaojun@jiangnan.edu.cnAbstract4040w=5,=0.5w=10,=0.5Wepresentanewlossfunction,namelyWingloss,forro-35w=5,=135w=10,=1w=5,=2w=10,=2bustfaciallandmarklocalisationwith

3、ConvolutionalNeu-30w=5,=330w=10,=3w=5,=4w=10,=4ralNetworks(CNNs).Wefirstcompareandanalysedif-2525ferentlossfunctionsincludingL2,L1andsmoothL1.The2020analysisoftheselossfunctionssuggeststhat,forthetrain-1515ingofaCNN-basedlocalisationmodel,moreattention1

4、010shouldbepaidtosmallandmediumrangeerrors.Tothisend,wedesignapiece-wiselossfunction.Thenewloss55amplifiestheimpactoferrorsfromtheinterval(-w,w)by00-20-1001020-20-1001020switchingfromL1losstoamodifiedlogarithmfunction.(a)w=5(b)w=10Toaddresstheproblemofun

5、der-representationofsam-Figure1.OurWinglossfunction(Eq.5)plottedwithdifferentpleswithlargeout-of-planeheadrotationsinthetrainingparametersettings,wherewlimitstherangeofthenon-linearpartset,weproposeasimplebuteffectiveboostingstrategy,re-andǫcontrolsthe

6、curvature.Bydesign,weamplifytheimpactofferredtoaspose-baseddatabalancing.Inparticular,wethesampleswithsmallandmediumrangeerrorstothenetworkdealwiththedataimbalanceproblembyduplicatingthetraining.minoritytrainingsamplesandperturbingthembyinject-ingrando

7、mimagerotation,boundingboxtranslationandotherdataaugmentationapproaches.Last,theproposedveryaccuratefaciallandmarklocalisationinconstrainedapproachisextendedtocreateatwo-stageframeworkforscenarios,evenusingtraditionalapproachessuchasAc-robustfacialland

8、marklocalisation.Theexperimentalre-tiveShapeModel(ASM)[7],ActiveAppearanceModelsultsobtainedonAFLWand300Wdemonstratethemerits(AAM)[8]andConstrainedLocalModel(CLM)[11].TheoftheWinglossfunction,andprovethesuperiorityoftheexistingchallenge