Structured-Forests-for-Fast-Edge-Detection

Seediscussions,stats,andauthorprofilesforthispublicationat:·December2013DOI:10.1109/ICCV.2013.231CITATIONS100READS3322authors,including:PiotrDollárFacebook50PUBLICATIONS4,379CITATIONSAllin-textreferencesunderlinedinbluearelinkedtopublicationsonResearchGate,lettingyouaccessandreadthemimmediately.Availablefrom:PiotrDollárRetrievedon:02May2016SEEPROFILE1StructuredForestsforFastEdgeDetectionPiotrDolla´[e-mail][e-mail]isacriticalcomponentofmanyvisionsystems,includingobjectdetectorsandimagesegmentationalgorithms.Patchesofedgesexhibitwell-knownformsoflocalstructure,suchasstraightlinesorT-junctions.Inthispaperwetakeadvantageofthestructurepresentinlocalimagepatchestolearnbothanaccurateandcomputation-allyefficientedgedetector.Weformulatetheproblemofpredictinglocaledgemasksinastructuredlearningframe-workappliedtorandomdecisionforests.Ournovelap-proachtolearningdecisiontreesrobustlymapsthestruc-turedlabelstoadiscretespaceonwhichstandardinfor-mationgainmeasuresmaybeevaluated.Theresultisanapproachthatobtainsrealtimeperformancethatisordersofmagnitudefasterthanmanycompetingstate-of-the-artapproaches,whilealsoachievingstate-of-the-artedgede-tectionresultsontheBSDS500SegmentationdatasetandNYUDepthdataset.Finally,weshowthepotentialofourapproachasageneralpurposeedgedetectorbyshowingourlearnededgemodelsgeneralizewellacrossdatasets.1.IntroductionEdgedetectionhasremainedafundamentaltaskincom-putervisionsincetheearly1970’s[13,10,32].Thedetec-tionofedgesisacriticalpreprocessingstepforavarietyoftasks,includingobjectrecognition[36,12],segmentation[23,1],andactivecontours[19].Traditionalapproachestoedgedetectionuseavarietyofmethodsforcomputingcolorgradientmagnitudesfollowedbynon-maximalsuppression[5,14,38].Unfortunately,manyvisuallysalientedgesdonotcorrespondtocolorgradients,suchastextureedges[24]andillusorycontours[29].State-of-the-artapproachestoedgedetection[1,31,21]useavarietyoffeaturesasinput,includingbrightness,colorandtexturegradientscomputedovermultiplescales.Fortopaccuracy,globalizationbasedonspectralclusteringmayalsobeperformed[1,31].Sincevisuallysalientedgescorrespondtoavarietyofvisualphenomena,findingaunifiedapproachtoedgede-tectionisdifficult.Motivatedbythisobservationseveralre-centpapershaveexploredtheuseoflearningtechniquesforedgedetection[9,37,21].EachoftheseapproachestakesFigure1.EdgedetectionresultsusingthreeversionsofourStruc-turedEdge(SE)detectordemonstratingtradeoffsinaccuracyvs.runtime.Weobtainrealtimeperformancewhilesimultaneouslyachievingstate-of-the-artresults.ODSnumberswerecomputedonBSDS[1]onwhichthehighlytunedgPbdetector[1]achievesascoreof.73.ThevariantsshownincludeSE-SS(T=1),SE-SS(T=4),andSE-MS,see§4fordetails.animagepatchandcomputesthelikelihoodthatthecenterpixelcontainsanedge.Theindependentedgepredictionsmaythenbecombinedusingglobalreasoning[37,1,31].Theedgesinalocalpatcharehighlyinterdependent[21].Theyoftencontainwell-knownpatterns,suchasstraightlines,parallellines,T-junctionsorY-junctions[30,21].Re-cently,afamilyoflearningapproachescalledstructuredlearning[26]hasbeenappliedtoproblemsexhibitingsim-ilarcharacteristics.Forinstance,[20]appliesstructuredlearningtotheproblemofsemanticimagelabelingforwhichlocalimagelabelsarealsohighlyinterdependent.Inthispaperweproposeageneralizedstructuredlearn-ingapproachthatweapplytoedgedetection.Thisap-proachallowsustotakeadvantageoftheinherentstructureinedgepatches,whilebeingsurprisinglycomputationallyefficient.Wecancomputeedgemapsinrealtime,whichisordersofmagnitudefasterthancompetingstate-of-the-artapproaches.Arandomforestframeworkisusedtocapturethestructuredinformation[20].Weformulatetheproblemofedgedetectionaspredictinglocalsegmentationmasksgiveninputimagepatches.Ournovelapproachtolearn-2ingdecisiontreesusesstructuredlabelstodeterminethesplittingfunctionateachbranchinthetree.Thestruc-turedlabelsarerobustlymappedtoadiscretespaceonwhichstandardinformationgainmeasuresmaybeevalu-ated.Eachforestpredictsapatchofedgepixellabelsthatareaggregatedacrosstheimagetocomputeourfinaledgemap,seeFigure1.Weshowstate-of-the-artresultsonboththeBSDS500[1]andtheNYUDepthdataset[33,31].Wedemonstratethepotentialofourapproachasageneralpur-poseedgedetectorbyshowingthestrongcrossdatasetgen-eralizationofourlearnededgemodels.1.1.RelatedworkInthissectionwediscussrelatedworkinedgedetectionandstructuredlearning.Edgedetection:Numerouspapershavebeenwrittenonedgedetectionoverthepast50years.Earlywork[13,10,5,27,14]focusedonthedetectionofintensityorcolorgradi-ents.ThepopularCannydetector[5]findsthepeakgradientmagnitudeorthogonaltotheedgedirection.Anevaluationofvariouslow-leveledgedetectorscanbefoundin[3]andanoverviewin[38].Morerecently,theworksof[24,22,1]exploreedgedetectioninthepresenceoftextures.Severaltechniqueshaveexploredtheuseoflearningforedgedetection[9,37,22,31,21].Dolla´retal.[9]usedaboostedclassifiert