Learning Hierarchical Features for Scene Labeling.pdf

《Learning Hierarchical Features for Scene Labeling.pdf》由會員上傳分享，免費在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。

1、1LearningHierarchicalFeaturesforSceneLabelingClementFarabet,CamilleCouprie,LaurentNajman,YannLeCun′AbstractScenelabelingconsistsinlabelingeachpixelinanimagewiththecategoryoftheobjectitbelongsto.Weproposeamethodthatusesamultiscaleconvolutionalnetworktrainedfromrawpixelstoextractdensefeaturevector

2、sthatencoderegionsofmultiplesizescenteredoneachpixel.Themethodalleviatestheneedforengineeredfeatures,andproducesapowerfulrepresentationthatcapturestexture,shapeandcontextualinformation.Wereportresultsusingmultiplepost-processingmethodstoproducethe?nallabeling.Amongthose,weproposeatechniquetoauto

3、maticallyretrieve,fromapoolofsegmentationcomponents,anoptimalsetofcomponentsthatbestexplainthescene;thesecomponentsarearbitrary,e.g.theycanbetakenfromasegmentationtree,orfromanyfamilyofover-segmentations.ThesystemyieldsrecordaccuraciesontheSiftFlowDataset(33classes)andtheBarcelonaDataset(170clas

4、ses)andnear-recordaccuracyonStanfordBackgroundDataset(8classes),whilebeinganorderofmagnitudefasterthancompetingapproaches,producinga320×240imagelabelinginlessthanasecond,includingfeatureextraction.IndexTermsConvolutionalnetworks,deeplearning,imagesegmentation,imageclassi?cation,sceneparsing.?1IN

5、TRODUCTIONthepresenceofahumanfacegenerallyindicatesthepresenceofahumanbodynearby),butmayalsodependMAGEUNDERSTANDINGisataskofprimaryimpor-onlong-rangeinformation.Forexample,identifyingaItanceforawiderangeofpracticalapplications.Onegreypixelasbelongingtoaroad,asidewalk,agraycar,importantsteptoward

6、sunderstandinganimageistoaconcretebuilding,oracloudyskyrequiresawidecon-performafull-scenelabelingalsoknownasasceneparsing,textualwindowthatshowsenoughofthesurroundingswhichconsistsinlabelingeverypixelintheimagetomakeaninformeddecision.Toaddressthisproblem,withthecategoryoftheobjectitbelongsto.A

7、fteraweproposetouseamulti-scaleconvolutionalnetwork,perfectsceneparsing,everyregionandeveryobjectiswhichcantakeintoaccountlargeinputwindows,whiledelineatedandtagged.Onechallengeofsceneparsingkeepingthenumberoffreeparameterst