Learning to Compare Image Patches via Convolutional Neural Networks

《Learning to Compare Image Patches via Convolutional Neural Networks》由會員上傳分享，免費在線閱讀，更多相關內(nèi)容在學術(shù)論文-天天文庫。

1、LearningtoCompareImagePatchesviaConvolutionalNeuralNetworksSergeyZagoruykoNikosKomodakisUniversiteParisEst,EcoledesPontsParisTechUniversiteParisEst,EcoledesPontsParisTechsergey.zagoruyko@imagine.enpc.frnikos.komodakis@enpc.frAbstractsimilarityInthispaperweshowhowtolearndirectlyfromimagedata(i.e.

2、,withoutresortingtomanually-designedfeatures)decisionnetworkageneralsimilarityfunctionforcomparingimagepatches,whichisataskoffundamentalimportanceformanycom-ConvNetputervisionproblems.Toencodesuchafunction,weoptforaCNN-basedmodelthatistrainedtoaccountforawidevarietyofchangesinimageappearance.Tot

3、hatend,weexploreandstudymultipleneuralnetworkarchitectures,whicharespeci?callyadaptedtothistask.Weshowthatsuchanapproachcansigni?cantlyoutperformthestate-of-patch1patch2the-artonseveralproblemsandbenchmarkdatasets.Figure1.Ourgoalistolearnageneralsimilarityfunctionforim-agepatches.Toencodesuchafu

4、nction,herewemakeuseofand1.Introductionexploreconvolutionalneuralnetworkarchitectures.Comparingpatchesacrossimagesisprobablyoneofthesoftware)largedatasetsthatcontainpatchcorrespondencesmostfundamentaltasksincomputervisionandimageanal-betweenimages[22].Thisbegsthefollowingquestion:canysis.Itisoft

5、enusedasasubroutinethatplaysanimportantwemakeproperuseofsuchdatasetstoautomaticallylearnroleinawidevarietyofvisiontasks.Thesecanrangefromasimilarityfunctionforimagepatches?low-leveltaskssuchasstructurefrommotion,widebaselineThegoalofthispaperistoaf?rmativelyaddressthematching,buildingpanoramas,a

6、ndimagesuper-resolution,abovequestion.Ouraimisthustobeabletogenerateauptohigher-leveltaskssuchasobjectrecognition,imagepatchsimilarityfunctionfromscratch,i.e.,withoutattempt-retrieval,andclassi?cationofobjectcategories,tomentioningtouseanymanuallydesignedfeaturesbutinsteaddi-afewcharacteristicex

7、amples.rectlylearnthisfunctionfromannotatedpairsofrawimageOfcourse,theproblemofdecidingiftwopatchescorre-patches.Tothatend,inspiredalsobytherecentadvancesinspondtoeachotherornotisquitechallengingasthereexistneuralarchitectur