[ICCV 2011] Tasting Families of Features for Image Classification

《[ICCV 2011] Tasting Families of Features for Image Classification》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、TastingFamiliesofFeaturesforImageClassificationCharlesDuboutandFranc¸oisFleuretIdiapResearchInstitutefcharles.dubout,francois.fleuretg@idiap.chAbstractinx5,vanillaBoostingofstumpsovermultiplefeaturesreachesstate-of-the-artperformance.However,suchtech-Usingmultiplefamiliesofimagefeaturesisaveryef-

2、niquesinducetwomajorpracticaldifficulties:thefirstistheficientstrategytoimproveperformanceinobjectdetectioncomputationalcostofthetraining,whichincreaseslinearlyorrecognition.However,suchastrategyinducesmultiplewiththenumberoffeatures,andthesecondisoverfittingchallengesformachinelearningmethods,bothf

3、romacom-thetrainingdata.Botharerelatedtothenumberoffeaturesputationalandastatisticalperspective.whichareactually“lookedat”duringtraining.ThemaincontributionofthispaperisanovelfeatureWeproposehereastraight-forwardandoriginalstrategysamplingproceduredubbed“Tasting”toimprovetheeffi-dubbed“Tasting”to

4、dealwiththatsituation,anduseittociencyofBoostinginsuchacontext.InsteadofsamplingimprovethelossreductioninBoosting.Thismethodsam-featuresinauniformmanner,Tastingcontinuouslyesti-plesafewfeaturesfromeveryfamilybeforethetrainingmatestheexpectedlossreductionforeachfamilyfromaperse,andstorestheirresp

5、onsesovereachtrainingsample.limitedsetoffeaturessampledpriortothelearning,andDuringBoosting,everytimewehavetosamplefeaturestobiasesthesamplingaccordingly.minimizeaweightederror,weusethesestoredfeaturestoWeevaluatetheperformanceofthisprocedurewithtensgetanestimateoftheexpectedreductionofthelossfo

6、reachoffamiliesoffeaturesonfourimageclassificationandob-family,andsampleaccordingly.jectdetectiondata-sets.WeshowthatTasting,whichdoesExperimentsonfourimageclassificationandobjectde-notrequirethetuningofanymeta-parameter,outperformstectionproblemsshowthatTastingsystematicallyoutper-systematicallyv

7、ariantsofuniformsamplingandstate-of-formssophisticatedbaselinesinminimizingboththetrain-the-artapproachesbasedonbanditstrategies.inglossandthetesterror,withoutrequiringthetuningofanyparameter,contrarilytothemostadvancedbasel