cvpr18-Weakly Supervised Coupled Networks for Visual Sentiment Analysis

《cvpr18-Weakly Supervised Coupled Networks for Visual Sentiment Analysis》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、WeaklySupervisedCoupledNetworksforVisualSentimentAnalysisJufengYang†,DongyuShe†,Yu-KunLai‡,PaulL.Rosin‡,Ming-HsuanYang§†CollegeofComputerandControlEngineering,NankaiUniversity,Tianjin,China‡SchoolofComputerScienceandInformatics,CardiffUniversity,Cardiff,UK§SchoolofEngineering,UniversityofCalifor

2、nia,Merced,USAAbstractAutomaticassessmentofsentimentfromvisualcontenthasgainedconsiderableattentionwiththeincreasingten-dencyofexpressingopinionson-line.Inthispaper,wesolvetheproblemofvisualsentimentanalysisusingthehigh-levelabstractionintherecognitionprocess.Existingmethod-sbasedonconvolutional

3、neuralnetworkslearnsentimentrepresentationsfromtheholisticimageappearance.How-ever,differentimageregionscanhaveadifferentinfluenceFigure1.ExamplesfromtheEmotionROIdataset[21].Thenor-malizedboundingboxesindicatetheregionsthatinfluencethee-ontheintendedexpression.Thispaperpresentsaweaklyvokedsentime

4、ntsannotatedby15users.Thefirsttwoexamplessupervisedcoupledconvolutionalnetworkwithtwobranch-arejoyimages,andthelasttwoexamplesaresadnessandfearim-estoleveragethelocalizedinformation.Thefirstbranchages,respectively.Ascanbeseen,thesentimentscanbeevokeddetectsasentimentspecificsoftmapbytrainingafullyc

5、on-byspecificregions.volutionalnetworkwiththecrossspatialpoolingstrategy,whichonlyrequiresimage-levellabels,therebysignificant-lyreducingtheannotationburden.Thesecondbranchu-addresssuchanabstracttaskasfollows.tilizesboththeholisticandlocalizedinformationbycou-First,visualsentimentanalysisismorecha

6、llengingthanplingthesentimentmapwithdeepfeaturesforrobustclas-conventionalrecognitiontasksduetoahigherlevelofsub-sification.Weintegratethesentimentdetectionandclassi-jectivityinthehumanrecognitionprocess[13].Itisnec-ficationbranchesintoaunifieddeepframeworkandopti-essarytotakemorecuesintoconsiderat

7、ionforvisualsen-mizethenetworkinanend-to-endmanner.Extensiveex-timentprediction.Figure1showsexamplesfromtheE-perimentsonsixbenchmarkdatasetsdemonstratethatthemotionROIdataset[21],whichprovidestheboundingboxproposedmethodperf