Segmentation-Aware Convolutional Networks Using Local Attention Masks

《Segmentation-Aware Convolutional Networks Using Local Attention Masks》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、Segmentation-AwareConvolutionalNetworksUsingLocalAttentionMasksAdamW.HarleyKonstantinosG.DerpanisIasonasKokkinosCarnegieMellonUniversityRyersonUniversityFacebookAIResearchaharley@cmu.edukosta@ryerson.caiasonask@fb.comAbstractWeintroduceanapproachtointegratesegmentationin-f

2、ormationwithinaconvolutionalneuralnetwork(CNN).Thiscounter-actsthetendencyofCNNstosmoothinfor-mationacrossregionsandincreasestheirspatialprecision.EmbedMaskEmbedMaskToobtainsegmentationinformation,wesetupaCNNtoprovideanembeddingspacewhereregionco-membershipcanbeestimatedba

3、sedonEuclideandistance.Weusetheseembeddingstocomputealocalattentionmaskrela-tivetoeveryneuronposition.WeincorporatesuchmasksinCNNsandreplacetheconvolutionoperationwithasegmentation-awarevariantthatallowsaneurontose-FilterFilterlectivelyattendtoinputscomingfromitsownregion.

4、Wecalltheresultingnetworkasegmentation-awareCNNbe-causeitadaptsitsfiltersateachimagepointaccordingtolocalsegmentationcues,whileatthesametimeremain-Normalizedfilterresponse≈Normalizedfilterresponseingfully-convolutional.WedemonstratethemeritofourFigure1:Segmentation-awareconvo

5、lutionfiltersareinvari-methodontwowidelydifferentdensepredictiontasks,thatanttobackgrounds.Weachievethisinthreesteps:(i)com-involveclassification(semanticsegmentation)andregres-putesegmentationcuesforeachpixel(i.e.,embeddings),sion(opticalflow).Ourresultsshowthatinsemanticseg

6、-(ii)createaforegroundmaskforeachpatch,and(iii)com-mentationwecanreplaceDenseCRFinferencewithacas-binethemaskswithconvolution,sothatthefiltersonlypro-cadeofsegmentation-awarefilters,andinopticalflowwecessthelocalforegroundineachimagepatch.obtainclearlysharperresponsesthantheo

7、nesobtainedwithcomparablenetworksthatdonotusesegmentation.Inbothcasessegmentation-awareconvolutionyieldssystem-dictionsthataresmoothandlow-resolution,resultingfromaticimprovementsoverstrongbaselines.therepeatedpoolingandsubsamplingstagesinthenet-workarchitecture,respective

8、ly.Thesestagesplayanim-portantroleinthehierarchicalconsolidationoffeatures,1.Introduction