(a)rawRGBimage(b)depth

image

(c)segmentation

labeling

(d)

foregroundness(e)

backgroundness(f)saliency

Fig.3.Overviewofoursaliencymappingalgorithm.RawRGB-Ddata(a-b)isusedtosegmentthesceneintohomogeneousregions,theresultingregionsareshownin(c).Inthenextstep,foregroundandbackgroundmeasuresarecomputedfromcolor,depthandthespatialdistributionoftheregions(d-e).The nalsaliencymapisobtainedbycombiningtheforegroundandbackgroundmaps(f).

regiontothecenteroftheimage.Itwillbelargeforobjectsbeingclosetocornersandsmallforobjectsaroundthecenteroftheimage.

TheBoundaryConnectivitymeasureproposedbyZhuetal.[9]isusedtoquantifyhowheavilyaregionriisconnectedtotheimageboundaries:

len(ri)

BCon(ri)=i(4)

asdepictedinFig.3(c).TwosamplesofforegroundandbackgroundsaliencymapsareshowninFig.3(d)andFig.3(e).Finally,theresultingsaliencymapisshowninFig.3(f).

III.EXPERIMENTALRESULT

Inthissection,weevaluateoursaliencydetectionmethodbyusingtwodifferentdatasets:theRGB-DdatasetprovidedbyCiptadietal.[4],whichweinthefollowingrefertoasDSD(depthsalientdata),andtheMSRAdatasetfromLiuetal.[19].Bothdatasetsincludeimageswithcomplexbackgroundandlowcontrastobjects,aswellasmanuallylabelledgroundtruthmasks(GT)forsalientobjectcandi-dates.Severalstate-of-the-artsaliencydetectionalgorithmsarechosenforcomparisonandareinthefollowingreferredtoasSF[3],MSS[14],IG[13],AC[18],IT[12],MZ[20]andSR[10],respectively.Tobemoreprecise,wecomparetheresultsofdifferentapproachesontheintroduceddatasetsifsaliencymapsareavailableforthecurrentbenchmarkdataset.Soforexample,RGBbasedalgorithmsarenotevaluatedontheRGB-Ddatasetsincetheyarenottailoredtomakeuseoftheadditionaldepthinformation(e.g.AC,SR).Ontheotherhand,weextendedsomeofthealgorithms,e.g.IG[13]toadditionallyutilizedepthinformationifthiswaspossibleinastraightforwardfashion.A.EvaluationonDSDdataset

TheDSDdatasetisanRGB-Ddataset,comprising80RGB-Dimagesusingamobilerobotinareal-worldindoorenvironment.Forperformanceevaluation,we rstgiveavisualcomparisonofdifferentmethodsonthisdataset.ThreeimagesampleresultsontheDSDdatasetareshowninthetopthreerowsofFig.4.The rstcolumnrepresentstheinputRGBimagesamplesandthelastcolumndepictsthebinarygroundtruthmasks.Visually,ourmethod(FBS)performsbestcomparedtoothermethodsinregardtotheGTmasksanddeliversbestresultsinregardtoourdesiredsaliency.Sinceperformanceofsaliencyishighlydependentonthedesiredpropertiesandtheinterpretationofimportantobjectsinscenes,itisgenerallynoteasytocomparedifferentmethods.Therefore,tobeabletoquantifythedifferentresultsoverthewholedatabaseandcomplyingwiththerelatedwork(e.g[3]),wechoosethemeanabsoluteerror(MAE)asameasure,whichsimplydescribesthedifferencebetweentheobtainedsaliencymapSandtheGT.

Conformingtothevisualimpression,Fig.5showsthatourmethodalsooutperformstheotherapproachesinregardto

wherelen(ri)isaregionperimeterontheboundary,area(ri)referstoitsarea.ThesalientregionhasasmallBCon(ri)value,comparedtothebackgroundregion.

Finally,wede neadissimilaritymeasureforbackgroundsaliencybetweenregionsas:BS(ri)=

N j=1

area(ri) area(rj) ·wbs(ri,rj)(5)

ThebackgroundsaliencyBS(ri)maybeinterpretedasthe

differenceofthescaleofregionricomparedtoallotherregionsrj.SimilarlytoFS,theGaussianweightwbs(ri,rj)isde nedas

(BCon(ri)+CDis(ri))2

(6)wbs(ri,rj)=1 exp δbswhicheffectivelydescribesthepositionoftheregion,with

valuesclosetozeroindicatingthattheregionisfarfromcornersandboundaries.Thebackgroundusuallyhasalargervaluethanotherregions.Theparameterδbscontrolstherangeofthebackgroundmeasure.D.FinalSaliencyMap

Becauseofthecomplementaritybetweenforegroundandbackgroundsaliency,wecombinethesetwokindsofsaliencymapstogetherwithdifferentweights.Wenormalizefore-groundFSaswellasbackgroundsaliencyBStotherange[0,1]andassumebothestimationstobeindependent.Hencewecombinetwomeasuresasfollowstocomputethe nalsaliencymapSal,

Sali=BSi·exp( t·FSi)

(7)

Theweighttisdeterminedaccordingtotheinformationcontainedinthecorrespondingmap.Inourcase,wesett=3asthescalingfactorthroughoutallexperiments.

AscanbeseeninFig.3,theinputRGB-Dimages(Fig.3(a)andFig.3(b))are rstmergedintohomogeneousregions