Recurrent Topic-Transition GAN for Visual Paragraph Generation

《Recurrent Topic-Transition GAN for Visual Paragraph Generation》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、RecurrentTopic-TransitionGANforVisualParagraphGenerationXiaodanLiangZhitingHuHaoZhangCarnegieMellonUniversityCarnegieMellonUniversityCarnegieMellonUniversityxiaodan1@cs.cmu.eduzhitingh@cs.cmu.eduhao@cs.cmu.eduChuangGanEricP.XingTsinghuaUniversityCarnegieMellonUniversityganch

2、uang1990@gmail.comepxing@cs.cmu.eduAbstracta)Genericdescription:AgroupofpeoplearesittingaroundalivingAnaturalimageusuallyconveysrichsemanticcontentroomtogether.Oneofthemeniswearingblacksleeveshirtandbluepants.Amanissittingandcanbeviewedfromdifferentangles.Existingim-nexttoth

3、ewoodentable.Amanandwomanaresittingonacouch.Thereisabrownagedescriptionmethodsarelargelyrestrictedbysmallwoodentableintheroom.setsofbiasedvisualparagraphannotations,andfailtocoverrichunderlyingsemantics.Inthispaper,weinves-b)Personalizeddescriptions:tigateasemi-supervisedpar

4、agraphgenerativeframeworkThereisamansittingonawoodenchair.Asmilingwomanissittingonacouch.ThemanwithawhiteremotewithwhiteShehasyellowshorthairandiswearingathatisabletosynthesizediverseandsemanticallycoher-buttonsiswearingablackandwhiteshortsleeveshirt.Sheisholdingawhiteentpar

5、agraphdescriptionsbyreasoningoverlocalseman-shirtandjeanpants.Awomannexttoplate.Thereisabrowncouchinthelivinghimhasredshirtsandredskirts.Thereroom.Infrontofherisawoodentable.ticregionsandexploitinglinguisticknowledge.Thepro-areamanandwomansittingonthefloorTherearepapersandgl

6、assesonthetable.nexttoawoodentable.posedRecurrentTopic-TransitionGenerativeAdversarialNetwork(RTT-GAN)buildsanadversarialframeworkbe-Figure1.OurRTT-GANisabletoautomaticallyproducegenerictweenastructuredparagraphgeneratorandmulti-levelparagraphdescriptionsshownin(a),andperson

7、alizeddescriptionsparagraphdiscriminators.Theparagraphgeneratorgen-bymanipulatingfirstsentences(highlightedinred),shownin(b).eratessentencesrecurrentlybyincorporatingregion-basedvisualandlanguageattentionmechanismsateachstep.datasets[22,34,17]pairingimageswithnaturallanguageT

8、hequalityofgeneratedparagraphsentencesisassesseddescriptions.Despitetheenco