Online-Object-Tracking-A-Benchmark

OnlineObjectTracking:[e-mail][e-mail][e-mail]gisoneofthemostimportantcomponentsinnumerousapplicationsofcomputervision.Whilemuchprogresshasbeenmadeinrecentyearswitheffortsonshar-ingcodeanddatasets,itisofgreatimportancetodevelopalibraryandbenchmarktogaugethestateoftheart.Afterbrieflyreviewingrecentadvancesofonlineobjecttracking,wecarryoutlargescaleexperimentswithvariousevalua-tioncriteriatounderstandhowthesealgorithmsperform.Thetestimagesequencesareannotatedwithdifferentat-tributesforperformanceevaluationandanalysis.Byana-lyzingquantitativeresults,weidentifyeffectiveapproachesforrobusttrackingandprovidepotentialfutureresearchdi-rectionsinthisfield.1.IntroductionObjecttrackingisoneofthemostimportantcomponentsinawiderangeofapplicationsincomputervision,suchassurveillance,humancomputerinteraction,andmedicalimaging[60,12].Giventheinitializedstate(e.g.,positionandsize)ofatargetobjectinaframeofavideo,thegoaloftrackingistoestimatethestatesofthetargetinthesub-sequentframes.Althoughobjecttrackinghasbeenstudiedforseveraldecades,andmuchprogresshasbeenmadeinre-centyears[28,16,47,5,40,26,19],itremainsaverychal-lengingproblem.Numerousfactorsaffecttheperformanceofatrackingalgorithm,suchasilluminationvariation,oc-clusion,aswellasbackgroundclutters,andthereexistsnosingletrackingapproachthatcansuccessfullyhandlealls-cenarios.Therefore,itiscrucialtoevaluatetheperformanceofstate-of-the-arttrackerstodemonstratetheirstrengthandweaknessandhelpidentifyfutureresearchdirectionsinthisfieldfordesigningmorerobustalgorithms.Forcomprehensiveperformanceevaluation,itiscriti-caltocollectarepresentativedataset.Thereexistsever-aldatasetsforvisualtrackinginthesurveillancescenarios,suchastheVIVID[13],CAVIAR[21],andPETSdatabas-es.However,thetargetobjectsareusuallyhumansorcarsofsmallsizeinthesesurveillancesequences,andtheback-groundisusuallystatic.Althoughsometrackingdataset-s[47,5,33]forgenericscenesareannotatedwithboundingbox,mostofthemarenot.Forsequenceswithoutlabeledgroundtruth,itisdifficulttoevaluatetrackingalgorithmsasthereportedresultsarebasedoninconsistentlyannotatedobjectlocations.Recently,moretrackingsourcecodeshavebeenmadepubliclyavailable,e.g.,theOAB[22],IVT[47],MIL[5],L1[40],andTLD[31]algorithms,whichhavebeencom-monlyusedforevaluation.However,theinputandoutputformatsofmosttrackersaredifferentandthusitisinconve-nientforlargescaleperformanceevaluation.Inthiswork,webuildacodelibrarythatincludesmostpubliclyavailabletrackersandatestdatasetwithground-truthannotationstofacilitatetheevaluationtask.Additionallyeachsequenceinthedatasetisannotatedwithattributesthatoftenaffecttrackingperformance,suchasocclusion,fastmotion,andilluminationvariation.Onecommonissueinassessingtrackingalgorithmsisthattheresultsarereportedbasedonjustafewsequenceswithdifferentinitialconditionsorparameters.Thus,theresultsdonotprovidetheholisticviewofthesealgorithm-s.Forfairandcomprehensiveperformanceevaluation,weproposetoperturbtheinitialstatespatiallyandtemporallyfromtheground-truthtargetlocations.Whiletherobust-nesstoinitializationisawell-knownprobleminthefield,itisseldomaddressedintheliterature.Tothebestofourknowledge,thisisthefirstcomprehensiveworktoaddressandanalyzetheinitializationproblemofobjecttracking.Weusetheprecisionplotsbasedonlocationerrormetricandthesuccessplotsbasedontheoverlapmetric,toana-lyzetheperformanceofeachalgorithm.Thecontributionofthisworkisthree-fold:Dataset.Webuildatrackingdatasetwith50fullyannotat-edsequencestofacilitatetrackingevaluation.Codelibrary.Weintegratemostpubliclyavailabletracker-sinourcodelibrarywithuniforminputandoutputformatstofacilitatelargescaleperformanceevaluation.Atpresent,itincludes29trackingalgorithms.Robustnessevaluation.Theinitialboundingboxesfortrackingaresampledspatiallyandtemporallytoevaluatetherobustnessandcharacteristicsoftrackers.Eachtrack-2013IEEEConferenceonComputerVisionandPatternRecognition1063-6919/13$26.00©2013IEEEDOI10.1109/CVPR.2013.31224092013IEEEConferenceonComputerVisionandPatternRecognition1063-6919/13$26.00©2013IEEEDOI10.1109/CVPR.2013.31224092013IEEEConferenceonComputerVisionandPatternRecognition1063-6919/13$26.00©2013IEEEDOI10.1109/CVPR.2013.3122411erisextensivelyevaluatedbyanalyzingmorethan660,000boundingboxoutputs.Thisworkmainlyfocusesontheonline1trackingofsin-gletarget.Thecodelibrary,annotateddatasetandallthetrackingresultsareavailableonthewebsite:targetrepresen-tationscheme,searchmechanism,andmodelupdate.Inaddition,somemethodshavebeenproposedthatbuildoncombingsometrackersorminingcontextinformation.RepresentationScheme.Objectrepresentationisoneofmajorcomponentsinanyvisualtrackerandnumerousschemeshavebeenpresented[35].Sincethepioneer-ingworkofLucasandKanade[37,8],holistictemplates(rawintensityvalues)havebeenwidelyusedfortrack