Robust computer vision Theory and applications

RobustComputerVision:TheoryandApplicationsNicuSebeMichaelS.LewLeidenInstituteofAdvancedComputerSciencePrefaceComputervisionistheenterpriseofautomatingandintegratingawiderangeofpro-cessesandrepresentationsusedforvisionperception.Itincludesmanytechniquesthatareusefulbythemselves,suchasimageprocessing(transforming,encoding,andtrans-mittingimages)andstatisticalpatternclassification(statisticaldecisiontheoryappliedtogeneralpatterns,visualorotherwise).Moreover,italsoincludestechniquesforge-ometricmodelingandcognitiveprocessing.Thefieldofcomputervisionmaybebestunderstoodbyconsideringdifferenttypesofapplications.Manyoftheseapplicationsinvolvetasksthatrequireeitherworkinahostileenvironment,ahighrateofprocessing,accessanduseoflargedatabasesofinformation,oraretediousforpeopletoperform.Computervisionsystemsareusedinmanyandvarioustypesofenvironments-frommanufacturingplants,tohospitalsurgicalsuits,andtothesurfaceofMars.Forexam-ple,inmanufacturingsystems,computervisionisoftenusedforqualitycontrol.Inthisapplication,thecomputervisionsystemscansmanufactureditemsfordefectsandpro-videscontrolsignalstoaroboticmanipulatortoremovedefectivepartsautomatically.Currentexamplesofmedicalsystemsbeingdevelopedinclude:systemstodiagnoseskintumorsautomatically,systemstoaidneurosurgeonsduringbrainsurgery,systemstoperformclinicaltestsautomatically,etc.ThefieldoflawenforcementandsecurityisalsoanactiveareaforcomputervisionsystemdevelopmentwithapplicationsrangingfromautomaticidentificationoffingerprintstoDNAanalysis.Inastandardapproach,statisticaltechniquesincomputervisionapplicationsmustestimateaccuratemodelparametersdespitesmall-scalenoiseinthedata,occasionallarge-scalemeasurementerrors(outliers),andmeasurementsfrommultiplepopulationsinthesamedataset.Increasingly,robustestimationtechniquesfromstatisticsarebe-ingusedtosolvetheseparameterestimationproblems.Ideally,thesetechniquesshouldeffectivelyignoretheoutlierswhenestimatingtheparametersofasinglepopulation.Inourapproach,weconsiderapplicationsthatinvolvesimilaritywherethegroundtruthisprovided.ThegoalistofindtheprobabilitydensityfunctionwhichmaximizestheVIIVIIIPrefacesimilarityprobability.Furthermore,wederivethecorrespondingmetricfromtheprob-abilitydensityfunctionbyusingthemaximumlikelihoodparadigmandweuseitintheexperiments.Thegoalofthisbookistodescribeandilluminatesomefundamentalprinciplesofrobustapproaches.Consequently,theintentionistointroducebasicconceptsandtechniquesofarobustapproachandtodevelopafoundation,whichcanbeusedinawidevarietyofcomputervisionalgorithms.Chapter1introducesthereadertotheparadigms,issues,andimportantapplicationsinvolvingvisualsimilarity,followedbyanin-depthchapter(Chapter2)whichdiscussesthemostinfluentialrobustframework-maximumlikelihood.Inrecentyears,thevisioncommunityhasgeneralizedbeyondgrayscalealgorithmstowardcolortechniqueswhichpromptsthethirdchapteroncolorbasedretrievalofimagesandobjects.Theotherprimaryfeatureswhicharefrequentlydiscussedinthevisionliteraturearetextureandshapewhicharecoveredinthefourthchapterandinthefifthchapter,respectively.Beyondclassificationalgorithms,thecomputervisionareahasbeeninterestedinfindingcorrespondencesbetweenpairsofimageswhichhavebeentakenfromdifferentspatialpositions(stereomatching)ordifferentmomentsintime(motiontracking).OuranalysisextendstobothofthesewithrespecttorecentdevelopmentsinrobusttechniquesinChapter5.Imagescontainingfacesareessentialtointelligentvision-basedhumancomputerin-teraction.Therapidlyexpandingresearchinfaceprocessingisbasedonthepremisethatinformationabouttheuser’sidentity,state,andintentcanbeextractedfromim-agesandthatcomputerscanthenreactaccordingly,e.g.,byobservingaperson’sfacialexpression.TheareaoffacialemotionrecognitioniscoveredinChapter7.Ineachofthechaptersweshowhowtheliteraturehasintroducedrobusttechniquesintotheparticulartopicarea,discusscomparativeexperimentsmadebyus,andcon-cludewithcommentsandrecommendations.Furthermore,wesurveythetopicareaanddescribetherepresentativeworkdone.Contents1Introduction11.1VisualSimilarity...............................21.1.1Color..................................51.1.2Texture................................71.1.3Shape.................................91.1.4Stereo.................................111.1.5Motion.................................131.1.6Facialexpression...........................141.1.7Summary...............................161.2EvaluationofComputerVisionAlgorithms.................171.3OverviewoftheBook.............................202MaximumLikelihoodFramework252.1Introduction..................................252.2StatisticalDistributions...........................272.2.1GaussianDistribution........................282.2.2ExponentialDistribution.......................392.2.3CauchyDistribution.........................422.3RobustStatistics...............................452.3.1Outliers................................462.4MaximumLikelihoodEstimators......................472.5MaximumLikelihoodinRelationtoOtherApproaches..........492.6OurMaximumLikelihoodApproach....................522.6