Efficient Retrieval of Deformable Shape Classes using Local Self-Similarities

《Efficient Retrieval of Deformable Shape Classes using Local Self-Similarities》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、EfficientRetrievalofDeformableShapeClassesusingLocalSelf-SimilaritiesKenChatfieldJamesPhilbinAndrewZissermanDept.ofEngineeringScienceDept.ofEngineeringScienceDept.ofEngineeringScienceUniversityofOxford,UKUniversityofOxford,UKUniversityofOxford,UKken.chatfield@oriel

2、.oxon.orgjames@robots.ox.ac.ukaz@robots.ox.ac.ukAbstractWepresentanefficientobjectretrievalsystembasedontheidentificationofabstractdeformable‘shape’classesusingtheself-similaritydescriptorofShechtmanandIrani[13].Givenauser-specifiedqueryobject,weretrieveFigure1:Chal

3、lengesofobjectclassidentification.Althoughotherimageswhichshareacommon‘shape’eveniftheirallfourimagesareofaheart,thereisnoobviousimagepropertyappearancediffersgreatlyintermsofcolour,texture,edges(e.g.texture,edgesorcolour)sharedbetweenthem.andothercommonphotometri

4、cproperties.Inordertousetheself-similaritydescriptorforeffi-cientretrievalwemakethreecontributions:(i)wespar-intensitygradients,arenotappropriate,andindeed(aswillsifythedescriptorpointsbylocatingdiscriminativeregionsbeshown)oftenperformpoorlyinsuchcases.withineach

5、image,thusreducingthecomputationalexpenseTheproblemofmatchingsuchshapesisaddressedbyofshapematching;(ii)weextend[13]toenablematch-thedescriptorofShechtmanandIrani[13],whichuseslo-ingdespitechangesinscale;and(iii)weshowthatvec-calself-similaritypatternsextractedfr

6、omtheimageasatorquantizingthedescriptordoesnotinhibitperformance,descriptor(reviewedinsection3).However,theirworkthusprovidingthebasisofalarge-scaleshape-basedre-concentratesonmatchingtemplatesatsimilarscalesandtrievalsystemusingabag-of-visual-wordsapproach.Per-d

7、oesnotaddresstheproblemsoffalsepositivematchesformanceisdemonstratedonthechallengingETHZde-orretrievalinlargedatasets.Thequestionsweinvestigateformableshapedatasetandafullepisodefromthetelevi-hereare:(i)isthedescriptorinvarianttochangesinscalesionseriesLost,andis

8、showntobesuperiortoappearance-and(ii)canitbeappliedtoefficientlargescaleimagere-basedapproachesformatchingnon-rigidshapeclasses.trieval[6,9,11,15]byvectorquanti