[IJCV 2010] Improving bag-of-features for large scale image search

《[IJCV 2010] Improving bag-of-features 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、Authormanuscript,publishedin"InternationalJournalofComputerVision87,3(2010)316-336"DOI:10.1007/s11263-009-0285-2Improvingbag-of-featuresforlargescaleimagesearchPre-printversionofJuly27th2009,correctedMarch15th2011HerveJ´egou´MatthijsDouzeCordeliaSchmidMarch15,2011Abstract

2、centimagesearchsystemsbuilduponthebag-of-featuresrepresentation,introducedinthecontextofimagesearchThisarticleimprovesrecentmethodsforlargescaleimagein[4].Descriptorsarequantizedintovisualwordswiththesearch.Wefirstanalyzethebag-of-featuresapproachinthek-meansalgorithm.Anima

3、geisthenrepresentedbytheframeworkofapproximatenearestneighborsearch.Thisfrequencyhistogramofvisualwordsobtainedbyassigningleadsustoderiveamorepreciserepresentationbasedoneachdescriptoroftheimagetotheclosestvisualword.Fast1)Hammingembedding(HE)and2)weakgeometriccon-accessto

4、thefrequencyvectorsisobtainedbyaninvertedsistencyconstraints(WGC).HEprovidesbinarysignaturesfilesystem.Notethatthisapproachisanapproximationthatrefinethematchingbasedonvisualwords.WGCfil-tothedirectmatchingofindividualdescriptorsandsome-tersmatchingdescriptorsthatarenotconsis

5、tentintermswhatdecreasesitsperformance.Itcomparesfavorablyinofangleandscale.HEandWGCareintegratedwithinantermsofmemoryusageagainstotherapproximatenearestinvertedfileandareefficientlyexploitedforallimagesinneighborsearchalgorithms,suchasthepopularEuclideanthedataset.Wethenint

6、roduceagraph-structuredquan-localitysensitivehashing(LSH)[7,8].LSHtypicallyre-tizerwhichsignificantlyspeedsuptheassignmentofthequires100–500bytesperdescriptortoindex,whichisnotdescriptorstovisualwords.Acomparisonwiththestateoftractable,asaonemillionimagedatasettypicallyprod

7、ucestheartshowstheinterestofourapproachwhenhighaccu-upto2billionlocaldescriptors.racyisneeded.SomerecentextensionsoftheBOFapproachspeedupExperimentsperformedonthreereferencedatasetsandtheassignmentofindividualdescriptorstovisualwords[5,adatasetofonemillionofimagesshowasign

8、ificantim-9]orthesearchforfrequencyvectors[10,11].Othersim-provementduetothebinarysignatur