Object Detection Networks on Convolutional Feature Maps

《Object Detection Networks on Convolutional Feature Maps》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、ObjectDetectionNetworksonConvolutionalFeatureMapsShaoqingRenKaimingHeRossGirshickXiangyuZhangJianSunMicrosoftResearchfv-shren,kahe,rbg,v-xiangz,jiansung@microsoft.comAbstractasaconvolutionalfeatureextractor,endingatthelastpool-inglayer,followedbyamulti-layerperceptron(MLP).ToMostobjectdet

2、ectorscontaintwoimportantcompo-date,evenwhenextratrainingdataareusedbytraditionalnents:afeatureextractorandanobjectclassifier.Themethods[34],theystilltrailfarbehinddeepConvNetsonfeatureextractorhasrapidlyevolvedwithsignificantre-detectionbenchmarks.searcheffortsleadingtobetterdeepConvNetarc

3、hitectures.Oneresearchstream[24,11,29,36]attemptingtobridgeTheobjectclassifier,however,hasnotreceivedmuchat-theperformancegapbetweentraditionaldetectorsanddeeptentionandmoststate-of-the-artsystems(likeR-CNN)useConvNetscreatesahybridofthetwo:thefeatureextractorsimplemulti-layerperceptrons.T

4、hispaperdemonstratesis“upgraded”toapre-traineddeepConvNet,buttheclassi-thatcarefullydesigningdeepnetworksforobjectclassifi-fierisleftasatraditionalmodel,suchasaDPM[24,11,29]cationisjustasimportant.Wetakeinspirationfromtradi-oraboostedclassifier[36].Thesehybridapproachesout-tionalobjectclassi

5、fiers,suchasDPM,andexperimentwithperformtheirHOG/SIFT/LBP-basedcounterparts[8,30],deepnetworksthathavepart-likefiltersandreasonoverbutstilllagfarbehindR-CNN,evenwhentheDPMislatentvariables.Wediscoverthatonpre-trainedconvolu-trainedend-to-endwithdeepConvNetfeatures[29].Inter-tionalfeaturemap

6、s,evenrandomlyinitializeddeepclassi-estingly,thedetectionaccuracyofthesehybridmethodsisfiersproduceexcellentresults,whiletheimprovementduetoclosetothatofR-CNNwhenusingalinearSVMonthelastfine-tuningissecondary;onHOGfeatures,deepclassifiersconvolutionalfeatures,withoutthefully-connectedlayers.

7、outperformDPMsandproducethebestHOG-onlyresultsTheSPPnetapproach[13]forobjectdetectionoccupieswithoutexternaldata.WebelievethesefindingsprovideamiddlegroundbetweenthehybridmodelsandR-CNN.newinsightfordevelopingobjectdetectionsystems.OurSPPnet,likethehybridmodelsbutunl