Object Detection
WeakHypothesesandBoostingforGeneric
ObjectDetectionandRecognition
A.Opelt1,2,M.Fussenegger1,2,A.Pinz2,andP.Auer1
InstituteofComputerScience,
8700Leoben,Austria
{auer,andreas.opelt}@unileoben.ac.at
InstituteofElectricalMeasurementandMeasurementSignalProcessing,
8010Graz,Austria
{fussenegger,opelt,pinz}@emt.tugraz.at12
Abstract.Inthispaperwedescribethe rststageofanewlearningsystemforobjectdetectionandrecognition.ForoursystemweproposeBoosting[5]astheunderlyinglearningtechnique.Thisallowstheuseofverydiversesetsof

visualfeaturesinthelearningprocesswithina

com-monframework:Boosting—togetherwithaweakhypotheses nder—maychooseveryinhomogeneousfeaturesasmostrelevantforcombina-tionintoa nalhypothesis.Asanotheradvantagetheweakhypotheses ndermaysearchtheweakhypothesesspacewithoutexplicitcalculationofallavailablehypotheses,reducingcomputationtime.ThiscontraststherelatedworkofAgarwalandRoth[1]whereWinnowwasusedaslearningalgorithmandallweakhypotheseswerecalculatedexplicitly.Inour rstempiricalevaluationweusefourtypesoflocaldescriptors:twobasiconesconsistingofasetofgrayvaluesandintensitymomentsandtwohighleveldescriptors:momentinvariants[8]andSIFTs[12].Thedescriptorsarecalculatedfromlocalpatchesdetectedbyaninter-estpointoperator.Theweakhypotheses nderselectsoneofthelocalpatchesandonetypeoflocaldescriptorande cientlysearchesforthemostdiscriminativesimilaritythreshold.Thisdi ersfromotherworkonBoostingforobjectrecognitionwheresimplerectangularhypotheses[22]orcomplexclassi ers[20]havebeenused.Inrelativelysimpleimages,wheretheobjectsareprominent,ourapproachyieldsresultscomparabletothestate-of-the-art[3].Butwealsoobtainverygoodresultsonmorecompleximages,wheretheobjectsarelocatedinarbitrarypositions,poses,andscalesintheimages.Theseresultsindicatethatour exibleapproach,whichalsoallowstheinclusionoffeaturesfromsegmentedre-gionsandevenspatialrelationships,leadsusasigni cantsteptowardsgenericobjectrecognition.
1Introduction
Webelievethatalearningcomponentisanecessarypartofanygenericob-jectrecognitionsystem.Inthispaperweinvestigateaprincipleapproachforlearningobjectsinstillimageswhichallowstheuseof exibleandextendibleT.PajdlaandJ.Matas(Eds.):ECCV2004,LNCS3022,pp.71–84,2004.
cSpringer-VerlagBerlinHeidelberg2004
7-Visual Object Tracking... 暂无评价 10页 2财富值 Face detection using ...knowledge can be used for the choice of the base hypotheses or weak ...
part hypotheses Grouping Grouping (object level) ... more complex detection 7 Object-centered, using ...G. Medioni, Generic shape learning and recognition...
ine tracking system for generic color objects. ...object trajectory tracking based on detection; 3)...feature and produce multiple hypotheses in every ...
ine tracking system for generic color objects. ...object trajectory tracking based on detection; 3)...feature and produce multiple hypotheses in every ...
热门文档