CopyRight 2012-2014 DS文库版权所有
ArcFace_Additive Angular Margin Loss for Deep Face Recognition
(0 次评价)1832 人阅读0 次下载

 ArcFace:AdditiveAngularMarginLossforDeepFaceRecognition JiankangDeng*ImperialCollegeLondon j.deng16@imperial.ac.uk JiaGuo ∗ InsightFace guojia@gmail.com NiannanXueImperialCollegeLondon n.xue15@imperial.ac.uk StefanosZafeiriouImperialCollegeLondon s.zafeiriou@imperial.ac.uk Abstract OneofthemainchallengesinfeaturelearningusingDeepConvolutionalNeuralNetworks(DCNNs)forlarge-scalefacerecognitionisthedesignofappropriatelossfunc-tionsthatenhancediscriminativepower.Centrelosspe-nalisesthedistancebetweenthedeepfeaturesandtheircor-respondingclasscentresintheEuclideanspacetoachieveintra-classcompactness.SphereFaceassumesthatthelin-eartransformationmatrixinthelastfullyconnectedlayercanbeusedasarepresentationoftheclasscentresinanangularspaceandpenalisestheanglesbetweenthedeepfeaturesandtheircorrespondingweightsinamultiplicativeway.Recently,apopularlineofresearchistoincorporatemarginsinwell-establishedlossfunctionsinordertomax-imisefaceclassseparability.Inthispaper,weproposeanAdditiveAngularMarginLoss(ArcFace)toobtainhighlydiscriminativefeaturesforfacerecognition.TheproposedArcFacehasacleargeometricinterpretationduetotheex-actcorrespondencetothegeodesicdistanceonthehyper-sphere.Wepresentarguablythemostextensiveexperimen-talevaluationofalltherecentstate-of-the-artfacerecog-nitionmethodsonover10facerecognitionbenchmarksin-cludinganewlarge-scaleimagedatabasewithtrillionlevelofpairsandalarge-scalevideodataset.WeshowthatAr-cFaceconsistentlyoutperformsthestate-of-the-artandcanbeeasilyimplementedwithnegligiblecomputationalover-head.Wereleaseallrefinedtrainingdata,trainingcodes,pre-trainedmodelsandtraininglogs 1 ,whichwillhelpre-producetheresultsinthispaper. 1.Introduction FacerepresentationusingDeepConvolutionalNeuralNetwork(DCNN)embeddingisthemethodofchoicefor ∗ denotesequalcontributiontothiswork. 1 https://github.com/deepinsight/insightface Figure1.Basedonthecentre[ 18 ]andfeature[ 37 ]normalisation,allidentitiesaredistributedonahypersphere.Toenhanceintra-classcompactnessandinter-classdiscrepancy,weconsiderfourkindsofGeodesicDistance(GDis)constraint.(A)Margin-Loss:insertageodesicdistancemarginbetweenthesampleandcen-tres.(B)Intra-Loss:decreasethegeodesicdistancebetweenthesampleandthecorrespondingcentre.(C)Inter-Loss:increasethegeodesicdistancebetweendifferentcentres.(D)Triplet-Loss:in-sertageodesicdistancemarginbetweentripletsamples.Inthispaper,weproposeanAdditiveAngularMarginLoss(ArcFace),whichisexactlycorrespondedtothegeodesicdistance(Arc)mar-ginpenaltyin(A),toenhancethediscriminativepoweroffacerecognitionmodel.Extensiveexperimentalresultsshowthatthestrategyof(A)ismosteffective. facerecognition[ 32 , 33 , 29 , 24 ].DCNNsmapthefaceim-age,typicallyafteraposenormalisationstep[ 45 ],intoafeaturethathassmallintra-classandlargeinter-classdis-tance.TherearetwomainlinesofresearchtotrainDCNNsforfacerecognition.Thosethattrainamulti-classclas-sifierwhichcanseparatedifferentidentitiesinthetrain-ingset,suchbyusingasoftmaxclassifier[ 33 , 24 , 6 ],andthosethatlearndirectlyanembedding,suchasthetripletloss[ 29 ].Basedonthelarge-scaletrainingdataandtheelaborateDCNNarchitectures,boththesoftmax-loss-basedmethods[ 6 ]andthetriplet-loss-basedmethods[ 29 ]canob-tainexcellentperformanceonfacerecognition.However, 1 arXiv:1801.07698v3 [cs.CV] 9 Feb 2019

打分:

0 星

用户评论:

小飞飞
于 2020-07-15 上传

版权及免责声明|RISC-V单片机中文网 |网站地图

GMT+8, 2022-10-2 13:49 , Processed in 1.056792 second(s), 30 queries .

返回顶部