2024年4月21日发(作者：塞班岛战役电影)

MIKOLAJCZYKANDSCHMID:APERFORMANCEEVALUATIONOFLOCALDESCRIPTORS1

Aperformanceevaluationoflocaldescriptors

KrystianMikolajczykandCordeliaSchmid

neeringScience

UniversityofOxford

Oxford,OX13PJ

UnitedKingdom

km@

INRIARhˆone-Alpes

655,’Europe

38330Montbonnot

France

schmid@

Abstract

Inthispaperwecomparetheperformanceofdescriptorscomputedforlocalinterestregions,asfor

exampleextractedbytheHarris-Affinedetector[32].Manydifferentdescriptorshavebeenproposedin

r,itisunclearwhichdescriptorsaremoreappropriateandhowtheirperformance

criptorsshouldbedistinctiveandatthesametimerobust

luationusesascriterion

recallwithreare

shapecontext[3],steerablefilters[12],PCA-SIFT[19],differentialinvariants[20],spinimages[21],

SIFT[26],complexfilters[37],momentinvariants[43],andcross-correlationfordifferenttypesof

proposeanextensionoftheSIFTdescriptor,andshowthatitoutperformsthe

rmore,weobservethattherankingofthedescriptorsismostlyindependentof

thesandsteerable

filtersshowthebestperformanceamongthelowdimensionaldescriptors.

IndexTerms

Localdescriptors,interestpoints,interestregions,invariance,matching,recognition.

I.I

NTRODUCTION

Localphotometricdescriptorscomputedforinterestregionshaveprovedtobeverysuccessful

inapplicationssuchaswidebaselinematching[37,42],objectrecognition[10,25],texture

jczyk,km@.

February23,2005DRAFT

MIKOLAJCZYKANDSCHMID:APERFORMANCEEVALUATIONOFLOCALDESCRIPTORS2

recognition[21],imageretrieval[29,38],robotlocalization[40],videodatamining[41],building

panoramas[4],andrecognitionofobjectcategories[8,9,22,35].Theyaredistinctive,robust

workhasconcentratedonmakingthese

aistodetectimageregionscovarianttoa

classoftransformations,whicharethenusedassupportregionstocomputeinvariantdescriptors.

Giveninvariantregiondetectors,theremainingquestionsarewhichisthemostappropriate

descriptortocharacterizetheregions,anddoesthechoiceofthedescriptordependontheregion

salargenumberofpossibledescriptorsandassociateddistancemeasureswhich

emphasizedifferentimagepropertieslikepixelintensities,color,texture,work

wefocusondescriptorscomputedongray-valueimages.

Theevaluationofthedescriptorsisperformedinthecontextofmatchingandrecognition

selecteda

numberofdescriptors,whichhavepreviouslyshownagoodperformanceinsuchacontextand

luationcriterion

isrecall-precision,r

possibleevaluationcriterionistheROC(ReceiverOperatingCharacteristics)inthecontextof

imageretrievalfromdatabases[6,31].Thedetectionrateisequivalenttorecallbutthefalse

posierefore

difficulttopredicttheactualnumberoffalsematchesforapairofsimilarimages.

Localfeatureswerealsosuccessfullyusedforobjectcategoryrecognitionandclassification.

Thecomparr,it

isunclearhowtoselectarepresentativesetofimagesforanobjectcategoryandhowtoprepare

thegroundtruth,sincethereisnolineartransformationrelatingimageswithinacategory.A

possiblesolutionistoselectmanuallyafewcorrespondingpointsandapplylooseconstraints

toverifycorrectmatches,asproposedin[18].

Inthispaperthecomparisoniscarriedoutfordifferentdescriptors,differentinterestregions

edtoourpreviouswork[31],thispaperperforms

ldescriptorsanddetectors

havebeenaddedtothecomparisonandthedatasetcontainsalargervarietyofscenestypes

modifiedtheevaluationcriterionandnowuserecall-precisionfor

kingofthetopdescriptorsisthesameasintheROCbasedevaluation[31].

February23,2005DRAFT