Object Detection Networks on Convolutional Feature Maps

《Object Detection Networks on Convolutional Feature Maps》由會員上傳分享，免費(fèi)在線閱讀，更多相關(guān)內(nèi)容在學(xué)術(shù)論文-天天文庫。

1、ObjectDetectionNetworksonConvolutionalFeatureMapsShaoqingRenKaimingHeRossGirshickXiangyuZhangJianSunMicrosoftResearchfv-shren,kahe,rbg,v-xiangz,jiansung@microsoft.comAbstractasaconvolutionalfeatureextractor,endingatthelastpool-inglayer,followedbyamulti-layerperceptron(MLP).ToMostobjectdet

2、ectorscontaintwoimportantcompo-date,evenwhenextratrainingdataareusedbytraditionalnents:afeatureextractorandanobjectclassi?er.Themethods[34],theystilltrailfarbehinddeepConvNetsonfeatureextractorhasrapidlyevolvedwithsigni?cantre-detectionbenchmarks.searcheffortsleadingtobetterdeepConvNetarc

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

4、hispaperdemonstratesis“upgraded”toapre-traineddeepConvNet,buttheclassi-thatcarefullydesigningdeepnetworksforobjectclassi?-?erisleftasatraditionalmodel,suchasaDPM[24,11,29]cationisjustasimportant.Wetakeinspirationfromtradi-oraboostedclassi?er[36].Thesehybridapproachesout-tionalobjectclassi

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

6、s,evenrandomlyinitializeddeepclassi-estingly,thedetectionaccuracyofthesehybridmethodsis?ersproduceexcellentresults,whiletheimprovementduetoclosetothatofR-CNNwhenusingalinearSVMonthelast?ne-tuningissecondary;onHOGfeatures,deepclassi?ersconvolutionalfeatures,withoutthefully-connectedlayers.

7、outperformDPMsandproducethebestHOG-onlyresultsTheSPPnetapproach[13]forobjectdetectionoccupieswithoutexternaldata.Webelievethese?ndingsprovideamiddlegroundbetweenthehybridmodelsandR-CNN.newinsightfordevelopingobjectdetectionsystems.OurSPPnet,likethehybridmodelsbutunl