Persistence of features for robust image matching

Frank Nielsen

Description

Niveau: Supérieur, Doctorat, Bac+8
Persistence of features for robust image matching Frank Nielsen Sylvain Boltz 05 mars 2010 1 Contexte Image matching is a central problem of image/video processing and com- puter vision. Classical methods go through two independent steps. First, they extract some geometric features in the image (typically from a corner or blob detection algorithm) and second, they match these features across two images. This approach raises two problems. On one hand, there is no gua- rantee that the detection algorithm will select the same geometric features in the second image. On the other hand, parts of image 1 can be invisible in image 2 and as a consequence, some features will be incorrectly matched. Figure 1 Matching features between images In order to tackle this stability issue, recent advancements in algebraic topology, and more precisely in persistence topology, have provided a mea- sure on the stability of the topology of a metric space under the perturbation of a filtering function [2]. This measure has already given promising results in clustering [1] and should be extended to the problem of image matching, jointly on the feature detection algorithm and on the matching algorithm. A survey of early applications in computer vision is a good starting point [3]. 1

• matching features


• persistence-based


• between image


• features across


• classical methods


• nical report


• image matching


• algebraic topology


• computational geometry

Sujets

• Nielsen
• Algebraic topology
• Computational geometry

Informations

• Publié par: profil-zyak-2012
• Publié le: 01 mars 2010
• Nombre de lectures: 40
• Langue: English
• Poids de l'ouvrage: 2 Mo

Extrait

PersistenceoffeaturesonearlythisforclusteringrobustMatcimagerematcofhingtlyFbrankimagesNielsentsSylvhaainofBoltzalready05tomarsthe2010o1hed.ConbtexterImagerecenmatcologyhingersistenceisvidedaofcenundertral[2].problemiofshouldimage/videoimageprodetectioncessinAgvisionandoincom-incorrectlyputer1vision.gClassicalwmethoorddstacgoythroughadvtalgebraicwandoinindepologyendenetmea-steps.stabilitFirst,toptheymetricextractpsomlteringemeasuregeometricenfeaturesresultsinantheeimageproblem(thing,ypicallythefromandahingcornereyorinblobadetectionstartingalgorith[3].me)matcandFiguresecond,theyhinmatcfeatureshettheseeenfeaturesInacrossettowkleostabilitimages.issue,Thistapproacancemenhinraisestopt,wmoopreciselyproblems.pOntopone,hand,vthereproisanosuregua-theranyteethethatologytheadetectionspacealgorithmthewillerturbationselectathefunctionsameThisgeometrichasfeaturesgivinpromthesingsecondinimage.[1]Ondthebotherextendedhand,thepartsofofmatcimagejoin1oncanfeaturebalgorithmeoninmatcvisiblealgorithm.insurvimageof2applicationsandcomputerasisagoconsequence,dsomepfeaturestwill1Figure2hGesPdiersiR?f?rencesstenSkraba.tH.ofer-homologies(onL.sublevOudot,elsetsinofINRIA,theher,funTctionDiscr2ainCadre.polytechnique.fr/~boltz/LeF.stageGuibd'unS.ePdur?ersistence-baseddemanifolds.2ort?2009.3D.moisA.send?roulologicaleandra&au,seinBoltzdehttp://www.lixl'?quip).e[1]deChazal,G?o-J.m?trieade,l'InformationY.duandLab.oratoirePd'Informatiqueclusteringderiemannianl'?coleResearcPRepolytec6968,h-Junenique,[2]sousEdelsbrunner,laLetscdirectionanddeZomoroFarank.Nielsenop(phttp://www.lixsistence.polytechnique.simplication.fr/~nielsen/ete)ComputationaletometrySylv228(4):511533,2002.LabTRep[3]computerD.h-FHPL2009375,reedman2009.andvision.C.ecChen.nicalAlgebraicorttopHPologyoratories,for3