cvpr18-bi directional message passing model for salient object用于显着对象检测双向消息传递模型.pdfVIP

ABi-directionalMessagePassingModelforSalientObjectDetection

11123

LuZhang,JuDai,HuchuanLu,YouHe,GangWang

1DalianUniversityofTechnology,

2NavalAviationUniversity,

3AlibabaAILabs,

{luzhangdut,daijucug}@,lhchuan@,

heyouf@126.com,wg134231@alibaba-.com

Recentprogressonsalientobjectdetectionisbenefi-

cialfromFullyConvolutionalNeurawork(FCN).The

saliencycuescontainedinmulti-levelconvolutionalfea-

turesarecomplementaryfordetectingsalientobjects.How

tointegratemulti-levelfeaturesbecomesanopenproblem

insaliencydetection.Inthispaper,weproposeanovel

bi-directionalmessagepassingmodeltointegratemulti-

levelfeaturesforsalientobjectdetection.Atfirst,weadopt

aMulti-scaleContext-awareFeatureExtractionModule(a)(b)(c)(d)

(MCFEM)formulti-levelfeaturemapstocapturerichcon-Figure1.VisualexamplesofourmethodandRFCN[29].From

textinformation.Thenabi-directionalstructureisde-lefttoright:(a)inputimage,(b)groundtruth,(c)saliencymapof

signedtopassmessagesbetweenmulti-levelfeatures,andtheproposedmethod,(d)saliencymapofRFCN[29].

agatefunctioniploitedtocontrolthemessagepassing

rate.Weusethefeaturesaftermessagepassing,whichsi-tion[31]andposeestimation[34].Motivatedbythis,sev-

multaneouslyencodesemanticinformationandspatialde-eralattemptshave