Instance-sensitive Fully Convolutional Networks英文学习资料

Instance-sensitive Fully Convolutional Networks英文学习资料

ID:40719152

大小:5.78 MB

页数:15页

时间:2019-08-06

Instance-sensitive Fully Convolutional Networks英文学习资料_第1页
Instance-sensitive Fully Convolutional Networks英文学习资料_第2页
Instance-sensitive Fully Convolutional Networks英文学习资料_第3页
Instance-sensitive Fully Convolutional Networks英文学习资料_第4页
Instance-sensitive Fully Convolutional Networks英文学习资料_第5页
资源描述:

《Instance-sensitive Fully Convolutional Networks英文学习资料》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、Instance-sensitiveFullyConvolutionalNetworksJifengDai1,KaimingHe1,YiLi2?,ShaoqingRen3?,JianSun112MicrosoftResearch,TsinghuaUniversity,3UniversityofScienceandTechnologyofChinaAbstractFullyconvolutionalnetworks(FCNs)havebeenprovenverysuccessfulforsemanticsegmentation,buttheFCNoutputsare

2、unawareofobjectinstances.Inthispaper,wedevelopFCNsthatarecapableofproposinginstance-levelsegmentcandidates.IncontrasttothepreviousFCNthatgeneratesonescoremap,ourFCNisdesignedtocomputeasmallsetofinstance-sensitivescoremaps,eachofwhichistheoutcomeofapixel-wiseclassifierofarelativepositio

3、ntoinstances.Ontopoftheseinstance-sensitivescoremaps,asimpleassemblingmoduleisabletooutputinstancecandidateateachposition.IncontrasttotherecentDeepMaskmethodforsegmentinginstances,ourmethoddoesnothaveanyhigh-dimensionallayerrelatedtothemaskresolution,butinsteadexploitsimagelocalcohere

4、nceforestimatinginstances.Wepresentcom-petitiveresultsofinstancesegmentproposalonbothPASCALVOCandMSCOCO.1IntroductionFullyconvolutionalnetworks(FCN)[1]havebeenprovenaneffectiveend-to-endsolutiontosemanticimagesegmentation.AnFCNproducesascoremapofasizeproportionaltotheinputimage,whereev

5、erypixelrepresentsaclassifierofobjects.Despitegoodaccuracyandeaseofusage,FCNsarenotdirectlyapplicableforproducinginstancesegments(Fig.1(top)).Previousinstancesemanticsegmentationmethods(e.g.,[2,3,4,5])ingeneralresortedtooff-the-shelfsegmentproposalmethods(e.g.,[6,7]).Inthispaper,wedevel

6、opanend-to-endfullyconvolutionalnetworkthatisarXiv:1603.08678v1[cs.CV]29Mar2016capableofsegmentingcandidateinstances.LiketheFCNin[1],inourmethodeverypixelstillrepresentsaclassifier;butunlikeanFCNthatgeneratesonescoremap(foroneobjectcategory),ourmethodcomputesasetofinstance-sensitivesco

7、remaps,whereeachpixelisaclassifierofrelativepositionstoanobjectinstance(Fig.1(bottom)).Forexample,witha3×3regulargriddepictingrelativepositions,weproduceasetof9scoremapsinwhich,e.g.,themap#6inFig.1hashighscoresonthe“rightside”ofobjectinstances.Withthissetofscoremaps,weareabletogenerate

8、anobj

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。