返回
LR-GAN Layered Recursive Generative Adversarial Networks for Image Generation

LR-GAN Layered Recursive Generative Adversarial Networks for Image Generation

ID：39715596

大小：4.59 MB

页数：21页

时间：2019-07-09

LR-GAN Layered Recursive Generative Adversarial Networks for Image Generation _第1页
LR-GAN Layered Recursive Generative Adversarial Networks for Image Generation _第2页
LR-GAN Layered Recursive Generative Adversarial Networks for Image Generation _第3页
LR-GAN Layered Recursive Generative Adversarial Networks for Image Generation _第4页
LR-GAN Layered Recursive Generative Adversarial Networks for Image Generation _第5页
LR-GAN Layered Recursive Generative Adversarial Networks for Image Generation _第6页
LR-GAN Layered Recursive Generative Adversarial Networks for Image Generation _第7页
LR-GAN Layered Recursive Generative Adversarial Networks for Image Generation _第8页
LR-GAN Layered Recursive Generative Adversarial Networks for Image Generation _第9页
LR-GAN Layered Recursive Generative Adversarial Networks for Image Generation _第10页
资源描述：

《LR-GAN Layered Recursive Generative Adversarial Networks for Image 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、PublishedasaconferencepaperatICLR2017LR-GAN:LAYEREDRECURSIVEGENERATIVEAD-VERSARIALNETWORKSFORIMAGEGENERATIONJianweiYangAnithaKannanDhruvBatraandDeviParikhVirginiaTechFacebookAIResearchGeorgiaInstituteofTechnologyBlacksburg,VAMenloPark,CAAtlanta,GAjw2yang@vt.eduakannan@fb.comfdbatra,par

2、ikhg@gatech.eduABSTRACTWepresentLR-GAN:anadversarialimagegenerationmodelwhichtakesscenestructureandcontextintoaccount.Unlikepreviousgenerativeadversarialnet-works(GANs),theproposedGANlearnstogenerateimagebackgroundandfore-groundsseparatelyandrecursively,andstitchtheforegroundsonthebackgro

3、undinacontextuallyrelevantmannertoproduceacompletenaturalimage.Foreachforeground,themodellearnstogenerateitsappearance,shapeandpose.Thewholemodelisunsupervised,andistrainedinanend-to-endmannerwithgra-dientdescentmethods.TheexperimentsdemonstratethatLR-GANcangeneratemorenaturalimageswithob

4、jectsthataremorehumanrecognizablethanDCGAN.1INTRODUCTIONGenerativeadversarialnetworks(GANs)(Goodfellowetal.,2014)haveshownsignificantpromiseasgenerativemodelsfornaturalimages.AflurryofrecentworkhasproposedimprovementsovertheoriginalGANworkforimagegeneration(Radfordetal.,2015;Dentonetal.,201

5、5;Salimansetal.,2016;Chenetal.,2016;Zhuetal.,2016;Zhaoetal.,2016),multi-stageimagegenerationincludingpart-basedmodels(Imetal.,2016;Kwak&Zhang,2016),imagegenerationconditionedoninputtextorattributes(Mansimovetal.,2015;Reedetal.,2016b;a),imagegenerationbasedon3Dstructure(Wang&Gupta,2016),an

6、devenvideogeneration(Vondricketal.,2016).Whiletheholistic‘gist’ofimagesgeneratedbytheseapproachesisbeginningtolooknatural,thereisclearlyalongwaytogo.Forinstance,theforegroundobjectsintheseimagestendtobedeformed,blendedintothebackground,andnotlookrealisticorrecognizable.Onefundamentallimit

7、ationofthesemethodsisthattheyattempttogenerateimageswithouttakingintoaccountthatimagesare2Dprojectionsofa3Dvisualworld,whichhasalotofstructuresinit.Thismanifestsasstructureinthe2Dimagesthatcapturethisworld.OneexampleofthisstructurearXiv:1703.01560v1[cs.CV]5Mar2017is

当前文档最多预览五页，下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页，下载文档查看全文
温馨提示：
1. 部分包含数学公式或PPT动画的文件，查看预览时可能会显示错乱或异常，文件下载后无此问题，请放心下载。
2. 本文档由用户上传，版权归属用户，天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容，确认文档内容符合您的需求后进行下载，若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误，付费完成后未能成功下载的用户请联系客服处理。