@[TMM 2010] Real-Time Visual Concept Classification

《@[TMM 2010] Real-Time Visual Concept 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、1Real-timeVisualConceptClassificationJ.R.R.Uijlings,A.W.M.SmeuldersandR.J.H.SchaAbstract—AsdatasetsgrowincreasinglylargeincontentbasedThispaperpresentsacomprehensiveevaluationofvariousimageandvideoretrieval,computationalefficiencyofconceptfastBag-of-Wordscomponentsintermsofbothcomputa-clas

2、sificationisimportant.Thispaperreviewstechniquestotionalefficiencyandretrievalperformance,whichisourmainaccelerateconceptclassification,whereweshowthetrade-offcontribution.Theevaluationshowsanincreaseofaccuracybetweencomputationalefficiencyandaccuracy.AsabasisweusetheBag-of-Wordsalgorithmtha

3、tinthe2008benchmarksofforRandomForests[7]bycombiningthemwithPrincipalTRECVIDandPASCALleadtothebestperformancescores.WeComponentAnalysis,andanincreaseincomputationaleffi-dividetheevaluationinthreesteps:(1)DescriptorExtraction,ciencybydeterminingrelevantimagedivisionsfortheSpatialwhereweeva

4、luateSIFT,SURF,DAISY,andSemanticTextons.Pyramid[8].Additionally,wepresentseveralimprovements(2)VisualWordAssignment,wherewecompareak-meansvisualuponthecomponentsunderconsideration:WeprovideavocabularywithaRandomForestandevaluatesubsampling,dimensionreductionwithPCA,anddivisionstrategieso

5、fthemodifiedwaytospeedupthecalculationofdenselysampled2SpatialPyramid.(3)Classification,whereweevaluatetheχ,SIFT[9]descriptors.Similarly,weturnSURF[10]intoRBF,andFastHistogramIntersectionkernelfortheSVM.afaster,denselysampleddescriptor.WeacceleratenearestApartfromtheevaluation,weaccelerate

6、thecalculationofneighbourassignment.Furthermore,weincreaseaccuracyofdenselysampledSIFTandSURF,acceleratenearestneighbourtheHistogramIntersectionbasedSupportVectorMachineassignment,andimproveaccuracyoftheHistogramIntersectionkernel.Weconcludebydiscussingwhetherfurtheraccelerationbybalanci

7、ngvisualwordfrequencies.Finally,nexttotheoftheBag-of-Wordspipelineispossible.experimentalevaluation,weprovideatheoreticaldiscussionOurresultsleadtoa7-foldspeedincreasewithoutaccuracyoncomputationalefficiencyoftheBag-of-Wordsmethod.loss,anda70-foldspeedincreasewith3%accurac