人工智能核心550篇论文 计算机视觉图像分割 1706.05587v3.pdfVIP

RethinkingAtrousConvolutionforSemanticImageSegmentation

Liang-ChiehChenGeorgePapandreouFlorianSchroffHartwigAdam

GoogleInc.

lcchen,gpapan,fschroff,hadam@

7AbstractConvConvConv

1kernel:3x3kernel:3x3kernel:3x3

0rate:1rate:6rate:24

2Inthiswork,werevisitatrousconvolution,apowerfultoolrate=24

toexplicitlyadjustfilter’sfield-of-viewaswellascontroltherate=6

crate=1

eresolutionoffeatureresponsescomputedbyDeepConvolu-

DtionalNeuralNetworks,intheapplicationofsemanticimage

segmentation.Tohandletheproblemofsegmentingobjects

5atmultiplescales,wedesignmoduleswhichemployatrous

]convolutionincascadeorinparalleltocapturemulti-scaleFeaturemapFeaturemapFeaturemap

Vcontextbyadoptingmultipleatrousrates.Furthermore,weFigure1.Atrousconvolutionwithkernelsizeanddifferent

CproposetoaugmentourpreviouslyproposedAtrousSpatialrates.Standardconvolutioncorrespondstoatrousconvolution

.PyramidPoolingmodule,whichprobesconvolutionalf