Improving Image Classification with Location Context

Improving Image Classification with Location Context

ID:40382090

大小:2.12 MB

页数:9页

时间:2019-08-01

Improving Image Classification with Location Context_第1页
Improving Image Classification with Location Context_第2页
Improving Image Classification with Location Context_第3页
Improving Image Classification with Location Context_第4页
Improving Image Classification with Location Context_第5页
资源描述:

《Improving Image Classification with Location Context》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、ImprovingImageClassificationwithLocationContextKevinTang1,ManoharPaluri2,LiFei-Fei1,RobFergus2,LubomirBourdev21ComputerScienceDepartment,StanfordUniversity2FacebookAIResearchfkdtang,feifeilig@cs.stanford.edufmano,robfergus,lubomirg@fb.comAbstractWiththewidespreada

2、vailabilityofcellphonesandcam-erasthathaveGPScapabilities,itiscommonforimagesbeinguploadedtotheInternettodaytohaveGPScoordi-natesassociatedwiththem.InadditiontoresearchthattriestopredictGPScoordinatesfromvisualfeatures,thisalsoopensupthedoortoproblemsthatarecondi

3、tionedon(a)(b)theavailabilityofGPScoordinates.Inthiswork,wetackletheproblemofperformingimageclassificationwithloca-tioncontext,inwhichwearegiventheGPScoordinatesforimagesinboththetrainandtestphases.Weexplorediffer-entwaysofencodingandextractingfeaturesfromtheGPSco

4、ordinates,andshowhowtonaturallyincorporatethesefeaturesintoaConvolutionalNeuralNetwork(CNN),thecurrentstate-of-the-artformostimageclassificationand(c)(d)recognitionproblems.WealsoshowhowitispossibletosimultaneouslylearntheoptimalpoolingradiiforasubsetFigure1.Which

5、oftheseareimagesofsnow?JustbylookingatofourfeatureswithintheCNNframework.Toevaluateourtheimages,itmaybedifficulttotell.However,whatifweknewmodelandtohelppromoteresearchinthisarea,weidentifythat(a)wastakenattheBonnevilleSaltFlatsinUtah,(b)wasasetoflocation-sensitiv

6、econceptsandannotateasubsetoftakeninNewHamsphire,(c)wastakeninDeathValley,CaliforniatheYahooFlickrCreativeCommons100Mdatasetthathasand(d)wastakennearPaloAlto,California?ImagecreditsgivenGPScoordinateswiththeseconcepts,whichwemakepub-insupplementarymaterial.liclya

7、vailable.Byleveraginglocationcontext,weareableareinterestedinclassifyingconsumerimageswithconceptstoachievealmosta7%gaininmeanaverageprecision.thatcommonlyoccurontheInternet,rangingfromobjectstoscenestospecificlandmarks,asthesearethethingsthatpeopleoftentakepictur

8、esof,andtheInternetisthelargest1.Introductionsourceofgeotaggedimages.BuildingontheCNNarchi-AsFigure1shows,itissometimeshardevenforhumanstectureintroducedin[26]

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。