CopyRight 2012-2014 DS文库版权所有
Fully Convolutional Networks for Semantic Segmentation
(0 次评价)1921 人阅读0 次下载

 FullyConvolutionalNetworksforSemanticSegmentation JonathanLong ∗ EvanShelhamer ∗ TrevorDarrellUCBerkeley { jonlong,shelhamer,trevor } @cs.berkeley.edu Abstract Convolutionalnetworksarepowerfulvisualmodelsthatyieldhierarchiesoffeatures.Weshowthatconvolu-tionalnetworksbythemselves,trainedend-to-end,pixels-to-pixels,exceedthestate-of-the-artinsemanticsegmen-tation.Ourkeyinsightistobuild“fullyconvolutional”networksthattakeinputofarbitrarysizeandproducecorrespondingly-sizedoutputwithefficientinferenceandlearning.Wedefineanddetailthespaceoffullyconvolu-tionalnetworks,explaintheirapplicationtospatiallydensepredictiontasks,anddrawconnectionstopriormodels.Weadaptcontemporaryclassificationnetworks(AlexNet[ 19 ],theVGGnet[ 31 ],andGoogLeNet[ 32 ])intofullyconvolu-tionalnetworksandtransfertheirlearnedrepresentationsbyfine-tuning[ 4 ]tothesegmentationtask.Wethende-fineanovelarchitecturethatcombinessemanticinforma-tionfromadeep,coarselayerwithappearanceinformationfromashallow,finelayertoproduceaccurateanddetailedsegmentations.Ourfullyconvolutionalnetworkachievesstate-of-the-artsegmentationofPASCALVOC(20%rela-tiveimprovementto62.2%meanIUon2012),NYUDv2,andSIFTFlow,whileinferencetakeslessthanonefifthofasecondforatypicalimage. 1.Introduction Convolutionalnetworksaredrivingadvancesinrecog-nition.Convnetsarenotonlyimprovingforwhole-imageclassification[ 19 , 31 , 32 ],butalsomakingprogressonlo-caltaskswithstructuredoutput.Theseincludeadvancesinboundingboxobjectdetection[ 29 , 12 , 17 ],partandkey-pointprediction[ 39 , 24 ],andlocalcorrespondence[ 24 , 9 ].Thenaturalnextstepintheprogressionfromcoarsetofineinferenceistomakeapredictionateverypixel.Priorapproacheshaveusedconvnetsforsemanticsegmentation[ 27 , 2 , 8 , 28 , 16 , 14 , 11 ],inwhicheachpixelislabeledwiththeclassofitsenclosingobjectorregion,butwithshort-comingsthatthisworkaddresses. ∗ Authorscontributedequally 96 38425640964096 2121 backward/learning forward/inference pixelwise prediction segmentation g.t. 256 384 Figure1.Fullyconvolutionalnetworkscanefficientlylearntomakedensepredictionsforper-pixeltaskslikesemanticsegmen-tation. Weshowthatafullyconvolutionalnetwork(FCN),trainedend-to-end,pixels-to-pixelsonsemanticsegmen-tationexceedsthestate-of-the-artwithoutfurthermachin-ery.Toourknowledge,thisisthefirstworktotrainFCNsend-to-end(1)forpixelwisepredictionand(2)fromsuper-visedpre-training.Fullyconvolutionalversionsofexistingnetworkspredictdenseoutputsfromarbitrary-sizedinputs.Bothlearningandinferenceareperformedwhole-image-at-a-timebydensefeedforwardcomputationandbackpropa-gation.In-networkupsamplinglayersenablepixelwisepre-dictionandlearninginnetswithsubsampledpooling.Thismethodisefficient,bothasymptoticallyandabso-lutely,andprecludestheneedforthecomplicationsinotherworks.Patchwisetrainingiscommon[ 27 , 2 , 8 , 28 , 11 ],butlackstheefficiencyoffullyconvolutionaltraining.Ourap-proachdoesnotmakeuseofpre-andpost-processingcom-plications,includingsuperpixels[ 8 , 16 ],proposals[ 16 , 14 ],orpost-hocrefinementbyrandomfieldsorlocalclassifiers[ 8 , 16 ].Ourmodeltransfersrecentsuccessinclassifica-tion[ 19 , 31 , 32 ]todensepredictionbyreinterpretingclas-sificationnetsasfullyconvolutionalandfine-tuningfromtheirlearnedrepresentations.Incontrast,previousworkshaveappliedsmallconvnetswithoutsupervisedpre-training[ 8 , 28 , 27 ].Semanticsegmentationfacesaninherenttensionbe-tweensemanticsandlocation:globalinformationresolveswhatwhilelocalinformationresolveswhere.Deepfeature 1 arXiv:1411.4038v2 [cs.CV] 8 Mar 2015

打分:

0 星

用户评论:

小飞飞
于 2020-07-15 上传

版权及免责声明|RISC-V单片机中文网 |网站地图

GMT+8, 2022-10-2 15:13 , Processed in 0.687241 second(s), 30 queries .

返回顶部