[CVPR 2013] Lp-norm IDF for Large Scale Image Search

《[CVPR 2013] Lp-norm IDF for Large Scale Image Search》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、2013IEEEConferenceonComputerVisionandPatternRecognitionLp-normIDFforLargeScaleImageSearchLiangZheng1,ShengjinWang1,ZiqiongLiu1,andQiTian21TsinghuaUniversity,Beijing,China2UniversityofTexasatSanAntonio,TX,78249,USAzheng-l06@mails.tsinghua.edu.cnwgsgj@tsinghua.edu.cnliuziqiong@ocrserv.ee.ts

2、inghua.edu.cnqitian@cs.utsa.eduAbstractTheInverseDocumentFrequency(IDF)isprevalentlyu-tilizedintheBag-of-Wordsbasedimagesearch.Thebasicideaistoassignlessweighttotermswithhighfrequency,andviceversa.However,theestimationofvisualwordfre-quencyiscoarseandheuristic.Therefore,theeffectiv

3、enessoftheconventionalIDFroutineismarginal,andfarfromoptimal.Totacklethisproblem,thispaperintroducesanov-elIDFexpressionbytheuseofLp-normpoolingtechnique.Carefullydesigned,theproposedIDFtakesintoaccounttheFigure1.Antoyexampleofanimagecollection.Visualwordstermfrequency,documentfrequency,t

4、hecomplexityofim-zxandzybothoccursinallthesiximages,butwithvaryingT-ages,aswellasthecodebookinformation.OptimizingtheFdistributionsovertheentireimagecollection.InconventionalIDFfunctiontowardsoptimalbalancingbetweenTFandIDF,theIDFweightsareequaltozeroforbothwords.ButwhenpIDFweightsyieldst

5、heso-calledLp-normIDF(pIDF).resortingtoTF,zxandzybothhavesomediscriminativepower,theproblemofwhichwillbetackledinthispaper.WeshowthattheconventionalIDFisaspecialcaseofourgeneralizedversion,andtwonovelIDFs,i.e.theaverageIDFandthemaxIDF,canalsobederivedfromourfor-mula.Further,bycountingfort

6、heterm-frequencyineachisthemostpopularandperhapsthemostsuccessfulone.image,theproposedLp-normIDFhelpstoalleviatethevi-Thismodelstartsfromtheextractionofsalientlocalre-sualwordburstinessphenomenon.gionsfromanimageandrepresentingeachlocalpatchasahigh-dimensionalfeaturevector(e.g.SIFT[7]orit

7、svariantsOurmethodisevaluatedthroughextensiveexperiments[13]).Thenthecontinuoushighdimensionalfeaturespaceonthreebenchmarkdatasets(Oxford5K,Paris6Kandisdividedintoadiscretespaceofvisualwords.ThisstepisFlickr1M).Wereportaperformanceimprovementofasachievedbyconstructi