AperformanceevaluationoflocaldescriptorsVIP专享

MIKOLAJCZYK AND SCHMID: A PERFORMANCE EVALUATION OF LOCAL DESCRIPTORS1A performanceevalu ationoflocaldescriptorsKry stianMikolajczy kandCordeliaSchmidDept.ofEngineeringScience INRIA Rhˆone-AlpesUniversityofOxford655,av.del’EuropeOxford,OX13PJ38330MontbonnotUnitedKingdomFrancekm@robots.ox.ac.ukschmid@inrialpes.frAbstractInthispaperwecomparetheperformanceofdescriptorscomputedforlocalinterestregions,asforexampleextractedby theHarris-Affinedetector[32].Manydifferentdescriptorshavebeenproposedintheliteratu re.However,itisu nclearwhichdescriptorsare moreappropriateandhowtheirperformancedependsontheinterestregiondetector.Thedescriptorsshou ldbe distinctiveandatthesametimerobu sttochangesinviewingconditionsaswellastoerrorsofthedetector.Ou r evalu ationu sesascriterionrecallwithrespecttoprecisionandiscarriedou tfordifferentimagetransformations.We compareshapecontext[3],steerablefilters[12],PCA-SIFT[19],differentialinvariants[20],spinimages[21],SIFT [26],complexfilters[37],momentinvariants[43],andcross-correlationfordifferentty pesofinterestregions.We alsoproposeanextensionoftheSIFT descriptor,andshowthatitou tperformstheoriginalmethod.Fu rthermore,weobservethattherankingofthedescriptorsismostlyindependentoftheinterestregiondetectorandthattheSIFT baseddescriptorsperformbest.Momentsandsteerablefiltersshowthebestperformanceamongthelowdimensionaldescriptors.Index TermsLocaldescriptors,interestpoints,interestregions,invariance,matching,recognition.I.INTRODUCTIONLocalphotometricdescriptorscompu tedforinterestregionshaveprovedtobe verysu ccessfu linapplicationssu chaswidebaselinematching[37,42],objectrecognition[10,25],textu reCorrespondingau thorisK.Mikolajczy k,km@robots.ox.ac.u k.Febru ary 23, 2005DRAFTMIKOLAJCZYK AND SCHMID: A PERFORMANCE EVALUATION OF LOCAL DESCRIPTORS2recognition[21],imageretrieval[29,38],robotlocalization[40],videodatamining[41],bu ildingpanoramas[4],andrecognitionofobjectcategories[8,9,22, 35].Th...