Robust Tracking via Convolutional Networks without Learning_20151014082723

Robust Tracking via Convolutional Networks without Learning_20151014082723

ID：40725156

大?。?.01 MB

頁數：13頁

時間：2019-08-06

Robust Tracking via Convolutional Networks without Learning_20151014082723_第1頁
Robust Tracking via Convolutional Networks without Learning_20151014082723_第2頁
Robust Tracking via Convolutional Networks without Learning_20151014082723_第3頁
Robust Tracking via Convolutional Networks without Learning_20151014082723_第4頁
Robust Tracking via Convolutional Networks without Learning_20151014082723_第5頁
資源描述：

《Robust Tracking via Convolutional Networks without Learning_20151014082723》由會員上傳分享，免費在線閱讀，更多相關內容在學術論文-天天文庫。

1、1RobustTrackingviaConvolutionalNetworkswithoutLearningKaihuaZhang,QingshanLiu,YiWu,andMing-HsuanYangAbstractDeepnetworkshavebeensuccessfullyappliedtovisualtrackingbylearningagenericrepresentationof?inefromnumeroustrainingimages.Howevertheof?inetrainin

2、gistime-consumingandthelearnedgenericrepresentationmaybelessdiscriminativefortrackingspeci?cobjects.Inthispaperwepresentthat,evenwithoutlearning,simpleconvolutionalnetworkscanbepowerfulenoughtodeveloparobustrepresentationforvisualtracking.Inthe?rstfra

3、me,werandomlyextractasetofnormalizedpatchesfromthetargetregionas?lters,whichde?neasetoffeaturemapsinthesubsequentframes.Thesemapsmeasuresimilaritiesbetweeneach?lterandtheusefullocalintensitypatternsacrossthetarget,therebyencodingitslocalstructuralinfo

4、rmation.Furthermore,allthemapsformtogetheraglobalrepresentation,whichmaintainstherelativegeometricpositionsofthelocalintensitypatterns,andhencetheinnergeometriclayoutofthetargetisalsowellpreserved.Asimpleandeffectiveonlinestrategyisadoptedtoupdatether

5、epresentation,allowingittorobustlyadapttotargetappearancevariations.Ourconvolutionnetworkshavesurprisinglylightweightstructure,yetperformfavorablyagainstseveralstate-of-the-artmethodsonalargebenchmarkdatasetwith50challengingvideos.IndexTermsVisualtrac

6、king,ConvolutionalNetworks,Deeplearning.arXiv:1501.04505v1[cs.CV]19Jan2015KaihuaZhang,QingshanLiuandYiWuarewithJiangsuKeyLaboratoryofBigDataAnalysisTechnology(B-DAT),NanjingUniversityofInformationScienceandTechnology.E-mail:fcskhzhang,qsliu,ywug@nuist

7、.edu.cn.Ming-HsuanYangiswithElectricalEngineeringandComputerScience,UniversityofCalifornia,Merced,CA,95344.E-mail:mhyang@ucmerced.edu.January20,2015DRAFT2Fig.1:Overviewoftheproposedrepresentation.Inputsamplesarewarpedintoacanonical3232images.We?rstra

8、ndomlyextractasetofnormalizedlocalpatchesfromthewarpedtargetregioninthe?rstframe,andthenusethemas?lterstoconvolveeachnormalizedsampleextractedfromsubsequentframes,resultinginasetoffeaturemaps.Finally,thefeaturemapsarestackedtogeneratethesample

當前文檔最多預覽五頁，下載文檔查看全文

此文檔下載收益歸作者所有

當前文檔最多預覽五頁，下載文檔查看全文
溫馨提示：
1. 部分包含數學公式或PPT動畫的文件，查看預覽時可能會顯示錯亂或異常，文件下載后無此問題，請放心下載。
2. 本文檔由用戶上傳，版權歸屬用戶，天天文庫負責整理代發(fā)布。如果您對本文檔版權有爭議請及時聯系客服。
3. 下載前請仔細閱讀文檔內容，確認文檔內容符合您的需求后進行下載，若出現內容與標題不符可向本站投訴處理。
4. 下載文檔時可能由于網絡波動等原因無法下載或下載錯誤，付費完成后未能成功下載的用戶請聯系客服處理。