Pixel-Level Matching for Video Object Segmentation Using Convolutional Neural Networks

《Pixel-Level Matching for Video Object Segmentation Using Convolutional Neural Networks》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、Pixel-LevelMatchingforVideoObjectSegmentationusingConvolutionalNeuralNetworksJaeShinYoon†‡FrancoisRameau‡JunsikKim‡SeokjuLee‡SeunghakShin‡InSoKweon‡†UMN,Minneapolis,MN‡KAIST,SouthKoreayoon0074@umn.edu,{frameau,Jskim2,sjlee,shshin}@rcv.kaist.ac.kr,iskweon@kaist.ac

2、.krAbstractSearchWeproposeanovelvideoobjectsegmentationalgorithmPixel-levelSimilaritybasedonpixel-levelmatchingusingConvolutionalNeuralFramet(Target)ObjectnessNetworks(CNN).Ournetworkaimstodistinguishthetar-getareafromthebackgroundonthebasisofthepixel-levelsimila

3、ritybetweentwoobjectunits.TheproposednetworkrepresentsatargetobjectusingfeaturesfromdifferentdepthQuerylayersinordertotakeadvantageofboththespatialdetailsResultandthecategory-levelsemanticinformation.Furthermore,Frame1(Reference)weproposeafeaturecompressiontechni

4、quethatdrasticallyreducesthememoryrequirementswhilemaintainingtheca-Figure1.Exampleofaresultoftheproposedpixel-levelmatchingpabilityoffeaturerepresentation.Two-stagetraining(pre-network.Here,acompletelysegmentedreferenceframeisfixedtrainingandfine-tuning)allowsourn

5、etworktohandleanytothequeryandasampletargetframeisfedtothesearchinput.targetobjectregardlessofitscategory(eveniftheobject’sEachcoloredboxisassociatedwiththesamecolorinFig.2.typedoesnotbelongtothepre-trainingdata)orofvaria-tionsinitsappearancethroughavideosequence

6、.Exper-Forinstance,fullyconnectedgraphshavebeenproposedinimentsonlargedatasetsdemonstratetheeffectivenessof[22]toconstructalongrangespatio-temporalgraphstruc-ourmodel-againstrelatedmethods-intermsofaccuracy,turerobusttochallengingsituationssuchasocclusion.Inspeed

7、,andstability.Finally,weintroducethetransferabil-anotherstudy[9],thehigherpotentialterminasupervoxelityofournetworktodifferentdomains,suchastheinfraredclusterunitwasusedtoenforcethesteadinessofagraphdatadomain.structure.Morerecently,non-localgraphconnectionsweree

8、ffectivelyapproximatedinthebilateralspace[17],which1.Introduction&RelatedWorksdrasticallyimprovedtheaccuracyofsegmentation.How-Videoobjectsegmentationreferstot