Salient Regions Detection for Indoor Robots using RGB-D Data(5)

(a)raw

RGB(b)IG

[13](c)MSS

[14](d)SF

[3]

(e)CP[4](top3)

andAC[18](bottom

3)

(f)

FBS(g)GT

Fig.4.Visualcomparisonofsaliencymapsobtainedbydifferentstate-of-the-artalgorithms.Ourmethodproducessaliencymapsthathighlightthewholeobjectregionandoutlinetheforegroundbetterthanothermethods.ThetopthreeimagesamplesaretakenfromtheRGB-Ddatasetpublishedin[4],thebottomthreerowsareRGBsampleswithoutdepthfromtheMSRAdatasetin

[19].

Fig.5.MeanabsoluteerrorsofdifferentalgorithmsontheDSDdataset.

MAE,andprovidesabetterestimateofthedesiredsaliencymaps.SincetheMAEisdesignedtocapturepixelintensitydifferences,itgivesagoodapproximationoftheperformanceofthecomparedalgorithms.

Nextweanalyzetheperformanceofouralgorithminregardtothedepthextensions.OurproposedapproachasexpecteddeliversthebestresultswhenweusethecombinedRGB-Dversion,althoughverygoodresultsare

already

Fig.6.MeanabsoluteerrorsonRGB-Ddataset(DSD[4]).Resultsobtainedwithourproposedmethod(FBS)usingcolor,depthandcombinedcoloranddepthinformation(left).AndcomparisontoalternativealgorithmsIG[13]extendedtousedepth(IGRGBD)andCP[4](right)

achievedwithcoloronlyascanbeseeninFig.6.FurthermorewecomparedtheIGalgorithm,thatweextendedtouseadditionaldepthinformation,withtheCPalgorithm,whichnaturallyistailoredfordepthdata.TheminorextensionstotheRGB-DversionoftheIGalgorithmmakeitalmostaspowerfulastheproposedmethod.