Robust Tracking via Convolutional Networks without Learning

《Robust Tracking via Convolutional Networks without Learning》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、1RobustTrackingviaConvolutionalNetworkswithoutLearningKaihuaZhang,QingshanLiu,YiWu,andMing-HsuanYangAbstractDeepnetworkshavebeensuccessfullyappliedtovisualtrackingbylearningagenericrepresentationofflinefromnumeroustrainingimages.Howevertheofflinetrainingistime-consumingandthelearnedgenericrepresentati

2、onmaybelessdiscriminativefortrackingspecificobjects.Inthispaperwepresentthat,evenwithoutlearning,simpleconvolutionalnetworkscanbepowerfulenoughtodeveloparobustrepresentationforvisualtracking.Inthefirstframe,werandomlyextractasetofnormalizedpatchesfromthetargetregionasfilters,whichdefineasetoffeaturemaps

3、inthesubsequentframes.Thesemapsmeasuresimilaritiesbetweeneachfilterandtheusefullocalintensitypatternsacrossthetarget,therebyencodingitslocalstructuralinformation.Furthermore,allthemapsformtogetheraglobalrepresentation,whichmaintainstherelativegeometricpositionsofthelocalintensitypatterns,andhencethei

4、nnergeometriclayoutofthetargetisalsowellpreserved.Asimpleandeffectiveonlinestrategyisadoptedtoupdatetherepresentation,allowingittorobustlyadapttotargetappearancevariations.Ourconvolutionnetworkshavesurprisinglylightweightstructure,yetperformfavorablyagainstseveralstate-of-the-artmethodsonalargebench

5、markdatasetwith50challengingvideos.IndexTermsVisualtracking,ConvolutionalNetworks,Deeplearning.arXiv:1501.04505v1[cs.CV]19Jan2015KaihuaZhang,QingshanLiuandYiWuarewithJiangsuKeyLaboratoryofBigDataAnalysisTechnology(B-DAT),NanjingUniversityofInformationScienceandTechnology.E-mail:fcskhzhang,qsliu,ywug

6、@nuist.edu.cn.Ming-HsuanYangiswithElectricalEngineeringandComputerScience,UniversityofCalifornia,Merced,CA,95344.E-mail:mhyang@ucmerced.edu.January20,2015DRAFT2Fig.1:Overviewoftheproposedrepresentation.Inputsamplesarewarpedintoacanonical3232images.Wefirstrandomlyextractasetofnormalizedlocalpatchesfr

7、omthewarpedtargetregioninthefirstframe,andthenusethemasfilterstoconvolveeachnormalizedsampleextractedfromsubsequentframes,resultinginasetoffeaturemaps.Finally,thefeaturemapsarestackedtogeneratethesample