[CVPR 2011] Unsupervised Auxiliary Visual Words Discovery for Large-Scale Image Object Retrieval

《[CVPR 2011] Unsupervised Auxiliary Visual Words Discovery for Large-Scale Image Object Retrieval》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、UnsupervisedAuxiliaryVisualWordsDiscoveryforLarge-ScaleImageObjectRetrievalYin-HsiKuo1;2,Hsuan-TienLin1,Wen-HuangCheng2,Yi-HsuanYang1,andWinstonH.Hsu11NationalTaiwanUniversityand2AcademiaSinica,Taipei,TaiwanAbstract
&'( 

2、 Imageobjectretrievallocatingimageoccurrencesofspecificobjectsinlarge-scaleimagecollectionsisessen-
 !
&tialformanipulatingthesheeramountofphotos.Cur-


3、  rentsolutions,mostlybasedonbags-of-wordsmodel,suf-ferfromlowrecallrateanddonotresistnoisescausedbythechangesinlighting,viewpoints,andevenocclusions.

!


4、"
#$
"Weproposetoaugmenteachimagewithauxiliaryvisualwords(AVWs),semanticallyrelevanttothesearchtargets.TheAVWsareautomaticallydiscoveredbyfeaturepropa-gationandselectionintextualandvisualimagegraphsin#
#""


5、
#%
#%$
!!#%anunsupervisedmanner.Weinvestigatevariantoptimiza-Figure1.Comparisonintheretrievalperformanceofthetradi-tionmethodsforeffectivenessandscalabilityinlarge-scaletionalBoWmodel[14]andtheproposedapproa

6、ch.(a)Anexam-imagecollections.Experimentinginthelarge-scalecon-pleofobject-levelqueryimage.(b)TheretrievalresultsofaBoWsumerphotos,wefoundthatthetheproposedmethodsig-model,whichgenerallysuffersfromthelowrecallrate.(c)Theresultsoftheproposedsystem,whichobta

7、insmoreaccurateandnificantlyimprovesthetraditionalbag-of-words(111%rel-diverseresults.Notethatthenumberbeloweachimageisitsrankatively).Meanwhile,theselectionprocesscanalsonotablyintheretrievalresultsandthenumberinaparenthesisrepresentsreducethenumberoffeatu

8、res(to1.4%)andcanfurtherfa-therankpredictedbytheBoWmodel.cilitateindexinginlarge-scaleimageobjectretrieval.DuetovariantcaptureconditionsandlargeVWvocabu-1.Introductionlary(e.g.,1millionvocabulary),thefeatures