[CVPR2015]Fully Convolutional Networks for Semantic Segmentation

《[CVPR2015]Fully Convolutional Networks for Semantic Segmentation》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、FullyConvolutionalNetworksforSemanticSegmentationJonathanLongEvanShelhamerTrevorDarrellUCBerkeleyfjonlong,shelhamer,trevorg@cs.berkeley.eduAbstractforward/inferencebackward/learningConvolutionalnetworksarepowerfulvisualmodelsthatsegmentationg.t.pixelwisepredictionyieldhierarchiesoffeatur

2、es.Weshowthatconvolu-tionalnetworksbythemselves,trainedend-to-end,pixels-21to-pixels,exceedthestate-of-the-artinsemanticsegmen-40964096384384256tation.Ourkeyinsightistobuild“fullyconvolutional”256networksthattakeinputofarbitrarysizeandproduce96correspondingly-sizedoutputwithefficientinferen

3、ceandlearning.Wedefineanddetailthespaceoffullyconvolu-21tionalnetworks,explaintheirapplicationtospatiallydenseFigure1.Fullyconvolutionalnetworkscanefficientlylearntomakedensepredictionsforper-pixeltaskslikesemanticsegmen-predictiontasks,anddrawconnectionstopriormodels.Wetation.adaptcontempor

4、aryclassificationnetworks(AlexNet[19],theVGGnet[31],andGoogLeNet[32])intofullyconvolu-tionalnetworksandtransfertheirlearnedrepresentationsWeshowthatafullyconvolutionalnetwork(FCN),byfine-tuning[4]tothesegmentationtask.Wethende-trainedend-to-end,pixels-to-pixelsonsemanticsegmen-fineanovelarchi

5、tecturethatcombinessemanticinforma-tationexceedsthestate-of-the-artwithoutfurthermachin-tionfromadeep,coarselayerwithappearanceinformationery.Toourknowledge,thisisthefirstworktotrainFCNsfromashallow,finelayertoproduceaccurateanddetailedend-to-end(1)forpixelwisepredictionand(2)fromsuper-segme

6、ntations.Ourfullyconvolutionalnetworkachievesvisedpre-training.Fullyconvolutionalversionsofexistingstate-of-the-artsegmentationofPASCALVOC(20%rela-networkspredictdenseoutputsfromarbitrary-sizedinputs.tiveimprovementto62.2%meanIUon2012),NYUDv2,Bothlearningandinferenceareperformedwhole-image

7、-at-andSIFTFlow,whileinferencetakeslessthanonefifthofaa-timebydensefeedforwardcomputationandbackpropa-secondforatypicalimage.gation.In-networkupsamplinglayersenablepixelwisepre-dictionandlearninginnetswithsubsampledpooling.Thismethodisefficient,bothasymptoticall