parisonofmeanabsoluteerrorsandprecision,recallandF-measurebarsfordifferentalgorithmsontheMSRA-1000RGBdatabase.

B.EvaluationonMSRAdataset

Tofurtherevaluateouralgorithm,wealsoperformex-perimentsontheMSRAdataset.MSRA[19]isastandardbenchmarkdatasetcomprising1000instanceswithaccuratehuman-labelledmasks(GT)forsalientobjects.Visualsam-plesareagainpresentedthebottomthreerowsinFig.4.Sincethisdatasetdoesnotprovidedepthimagesamples,wepluginaSLICbasedsegmentationintotheproposedmethodwhichdoesnotdependondepthdata.Fortheotherstagesofourpipeline,wecreatedummydepthimageswithalldepthvaluesinitializedtozero.Furthermore,weevaluateandcomparetotheRGB-onlybasedACalgorithminsteadofthepreviouslyevaluatedCPalgorithmonRGB-D.Thevisionsamplesagaindemonstrate,thatoursaliencyapproachisabletocompetewithotherstateoftheartmethods,evenwhenonlycolorisavailable.Inour nalanalysisweadditionallyevaluatedprecisionandrecallbarsonthewholeMSRAdatasetasdepictedinFig.7.Theobtainedresultsonceagainemphasizetherobustnessandaccuracyofourproposedmethod,evenoncolor-onlyimagesamples.

IV.CONCLUSION

Inthispaper,weproposedasaliencydetectionmethodwhichisabletosuccessfullydetectmultipleobjectsonamobilerobot.Tosolvethede nedtaskofdetectingsalientobjectsinthescene,weproposeanimprovedRGB-Dbasedsegmentationmethodtobeabletoautomaticallyclustermajorregionsofinterest.Theclusteringmaintainstheuniquenessandconsistencyofdifferentobjectsanddecreasescomputationalcomplexityforthelaterstages,whichrelyoncoloranddepthinformation.Furthermore,weproposedanewmethodtoeffectivelygenerateasaliencymapbycombiningforegroundandbackgroundsaliencymaps.Weutilizedcontrast-basedsaliencymeasuresbycolor,depthandspatiallayoutinformation.Depthwasusedtorecoupcomplementaryobjectinformationforsegmentationaswellassaliencytasks.Theproposedapproachisverywellsuitedtocompetewithstate-of-the-artsaliencyalgorithmsandevenoutperformedthemonbothevaluateddatasets.

REFERENCES

[1]P.Wang,L.Zhou,G.Wei,andH.Qiao,“salientregiondetection

beasedonlocalandblobalsaliency,”inIEEEInt.Conf.RoboticsandAutomation(ICRA),(Hongkong,China),pp.1546–1551,May2014.

[2]M.-M.Cheng,N.J.Mitra,X.Huang,P.H.S.Torr,andS.-M.Hu,

“Globalcontrastbasedsalientregiondetection,”IEEETrans.PatternAnalysisandMachineIntelligence(PAMI),pp.433–449,2014.[3]F.Perazzi,P.Kr¨ahenb¨uhl,Y.Pritch,andA.Hornung,“Saliency lters:

Contrastbased lteringforsalientregiondetection,”puterVisionandPatternRecognition(CVPR),pp.733–740,2012.[4]A.Ciptadi,T.Hermans,andJ.M.Rehg,“AnInDepthViewof

Saliency,”inBritishMachineVisionConference(BMVC),(Bristol,UnitedKingdom),September2013.

[5]ng,T.V.Nguyen,H.Katti,K.Yadati,M.S.Kankanhalli,and

S.Yan,“Depthmatters:In uenceofdepthcuesonvisualsaliency,”puterVision(ECCV),vol.7573,(Berlin,Heidelberg),pp.101–115,2012.

[6]L.Jiang,A.Koch,S.A.Scherer,andA.Zell,“Multi-classfruit

classi cationusingRGB-Ddataforindoorrobots,”inIEEEInt.Conf.RoboticsandBiomimetics(ROBIO),(Shenzhen),Dec.2013.

[7]L.Jiang,A.Koch,andA.Zell,“Objectrecognitionandtracking

forindoorrobotsusinganrgb-dsensor,”inInt.Conf.IntelligentAutonomousSystems(IAS-13),(Padova,Italy),Jul.2014.

[8]H.Jiang,J.Wang,Z.Yuan,Y.Wu,N.Zheng,andS.Li,“Salientobject

detection:Adiscriminativeregionalfeatureintegrationapproach,”puterVisionandPatternRecognition(CVPR),(Portland),pp.2083–2090,Jun.2013.

[9]W.Zhu,S.Liang,Y.W.Wei,andJ.Sun,“Saliencyoptimizationfrom

robustbackgrounddetection,”puterVisionandPatternRecognition(CVPR),(Columbus,Ohio),Jun.2014.

[10]X.HouandL.Zhang,“Saliencydetection:Aspectralresidualap-proach,”puterVisionandPatternRecognition(CVPR),pp.1–8,2007.

[11]L.Mai,Y.Niu,andF.Liu,“Saliencyaggregation:Adata-drivenap-proach,”puterVisionandPatternRecognition(CVPR),pp.1131–1138,June2013.

[12]L.Itti,C.Koch,andE.Niebur,“Amodelofsaliency-basedvisual

attentionforrapidsceneanalysis,”IEEETrans.PatternAnalysisandMachineIntelligence(PAMI),vol.20,pp.1254–1259,nov.1998.[13]R.Achanta,S.S.Hemami,F.J.Estrada,andS.Ssstrunk,“Frequency-tunedsalientregiondetection,”puterVisionandPatternRecognition(CVPR),pp.1597–1604,2009.

[14]R.AchantaandS.Ssstrunk,“SaliencyDetectionusingMaximum

SymmetricSurround,”inIEEEInt.Conf.ImageProcessing(ICIP),(HongKong),Sep.2010.

[15]A.Maki,P.Nordlund,andJ.-O.Eklundh,“Acomputationalmodelof

depth-basedattention,”inIEEEInt.Conf.PatternRecognition(ICPR),vol.4,pp.734–739,Aug1996.

[16]N.OuerhaniandH.Hugli,“Computingvisualattentionfromscene

depth,”inIEEEInt.Conf.PatternRecognition(ICPR),vol.1,pp.375–378vol.1,Sep.2000.

[17]P.F.FelzenszwalbandD.P.Huttenlocher,“Ef cientgraphbased

imagesegmentation,”put.Vision,vol.59,pp.167–181,sep.2004.

[18]R.Achanta,F.Estrada,P.Wils,andS.S¨usstrunk,“Salientregion

detectionandsegmentation,”puterVisionSystems,(Berlin,Heidelberg),pp.66–75,2008.

[19]T.Liu,Z.Yuan,J.Sun,J.Wang,N.Zheng,X.Tang,andH.-Y.Shum,

“Learningtodetectasalientobject,”IEEETrans.PatternAnalysisandMachineIntelligence(PAMI),vol.33,pp.353–367,feb.2011.

[20]Y.-F.MaandH.-J.Zhang,“Contrast-basedimageattentionanalysis

byusingfuzzygrowing,”inPro.Conf.Multimedia,(NewYork,NY,USA),pp.374–381,ACM,2003.