So, you found all these nice features…what is the problem?Visual areas in the brainVisual areas in the brainImage removed for copyright reasons.Image removed for copyright reasons.CORNEACORNEAAQUEOUSAQUEOUSHUMORHUMORFOVEAFOVEAOPTICOPTICNERVENERVEIRISIRISPUPILPUPILLENSLENSVITREOUSVITREOUSHUMORHUMORRETINARETINAWhat do you see?Why?The world is a complicated placeThe world is a complicated placeImage removed for copyright reasons.Image removed for copyright reasons.CORNEACORNEAAQUEOUSAQUEOUSHUMORHUMORFOVEAFOVEAOPTICOPTICNERVENERVEIRISIRISPUPILPUPILLENSLENSVITREOUSVITREOUSHUMORHUMORRETINARETINAhttp://web.mit.edu/bcs/schillerlab/research/A-Vision/A2-1.htmlCourtesy of Peter Schiller. Used with permission.Outside viewOutside viewVisual CortexVisual CortexView from theView from themiddleImage removed for copyright reasons.middleFlatten the brainFlatten the brain(like making a map out of a globe,(like making a map out of a globe,Only worse)Only worse)Image removed for copyright reasons.Do we really have centerDo we really have center--surround receptive fields?surround receptive fields?The Hermann GridThe Hermann GridDo we really have centerDo we really have center--surround receptive fields?surround receptive fields?The Hermann GridThe Hermann GridDo we really have centerDo we really have center--surround receptive fields?surround receptive fields?Umm…what is Umm…what is happening here?happening here?Bright LightDim LightBright LightDim LightTimeThe stimulusA neuron's responseLight% "I see it"Responses1007550250you wantyou getColorColorHow do you see color?How do you see color?Wavelength (nm)400 500 600 700Wavelength 1 produces a response of size XWavelength 1 produces a response of size XWavelength (nm)400 500 600 700Wavelength 2Wavelength 2produces a response of size Xproduces a response of size XWavelength (nm)400 500 600 700Two wavelengths,one response.The problem of “The problem of “univarianceunivariance””So, we have a problem.So, we have a problem.Wavelength (nm)400 500 600 700Here is a solution….add another cone type. Here is a solution….add another cone type. Wavelength (nm)400 500 600 700Two cones can give you color visionTwo cones can give you color visionWavelength (nm)400 500 600 700XYZX/Y = X/Y = redred, X/Z = , X/Z = greengreenCOMPARISONS ARE CRITICALCOMPARISONS ARE CRITICALThree cones give you Three cones give you TrichromacyTrichromacyWavelength (nm)400 500 600 700LMSThree cones give you Three cones give you TrichromacyTrichromacyWavelength (nm)400 500 600 700LMSAnyAnylight = light = aLaL+ + bMbM+ + cScSLet’s add some patches togetherLet’s add some patches togetherWavelength (nm)Let’s take GREENLet’s take GREENWavelength (nm)M1L1And add REDAnd add REDWavelength (nm)M1L1M2L2Red + GreenRed + Green= (M1+M2)/(L1+L2) = 1= (M1+M2)/(L1+L2) = 1Compare that to YELLOWCompare that to YELLOWWavelength (nm)M3L3YellowYellow= M3/L3 = 1= M3/L3 = 1It follows thatIt follows thatREDREDIt follows thatIt follows thatplusplusREDREDGREENGREENYields Yields YellowYellowR+G and Y areR+G and Y areMETAMERSMETAMERSThis is ADDITIVE color mixtureThis is ADDITIVE color mixtureBut what about color paint in kindergarten?But what about color paint in kindergarten?x xxxx x xxxxxxxx x xxxBlue paintBlue paintYellow paintYellow paintMixing paint is SUBTRACTIVEMixing paint is SUBTRACTIVEx xxxxxxxxThe The intersection intersection ofofBlue paint and Yellow Blue paint and Yellow paint looks paint looks GreenGreenRecall…Three cones give you Recall…Three cones give you TrichromacyTrichromacyWavelength (nm)400 500 600 700LMSSuppose: Suppose: if S=M=L, then WHITEif S=M=L, then WHITESuppose that L gets Suppose that L gets tired?tired?Wavelength (nm)400 500 600 700LMSWhat does S=M>L look like?What does S=M>L look like?Pretty boring……Pretty boring……Pretty ……Pretty ……Pretty, not boring……Pretty, not boring……Try thisTry thisNegative afterimageNegative afterimageVertical and Horizontal look the same?Vertical and Horizontal look the same?Go forwardGo forwardGo backGo backVertical and Horizontal look the same?Vertical and Horizontal look the same?So, you found all these nice features…what is the problem?Which lines group together?Which lines group together?How about here? Why?How about here? Why?Which gray line is a likelycontinuation of the blackline?good betnot badmaybenot likelygood betnot badWhich gray line is a likelycontinuation of the blackline?good betnot badmaybenot likelyWHAT IS THIS?Does this seem likely?This seems more likely“Good continuation”One curved line or three?You ‘know’ about occlusionOne curved line or three?You ‘know’ about occlusionOrganized by columns or rows? Organized by columns or rows?Now? Organized by columns or rows? Why? Now? Organized by columns or rows? Why? ProximityProximityNow? Organized by columns or rows? Why? Now? Organized by columns or rows? Why? Did Similarity trump Proximity?Did Similarity trump Proximity?Let’s magnify the critical bit.See that rectangle?See that rectangle?How about that rectangle?How about that rectangle?How about that circle?How about that circle?Not as good?Not as good?Edges are importantThe visual system distinguishes “real” edges from shadowsThe visual system distinguishes “real” edges from shadowsImage removed for copyright reasons.Remember: You want to know about the world, not your retinaRemember: You want to know about the world, not your retinaCOWCOWMinimal shadow can give you facesImages removed for copyright reasons.Faces from University of Bielefeld Cognitive Robot project.Depth CuesFrom 2D-3DImage removed for copyright reasons.OcclusionIs this likely?SizeTextureRelative position (height in field)Here is why it worksYou don’t need to recognize the objectsAreal Perspective (haze)The misty mountains far awayLinear PerspectiveVanishing pointLinear Perspective?Image removed for copyright reasons.Where is the vanishing point?Linear Perspective?These local bits don’t add upLinear Perspective?These add up…ambiguouslyShadowsShadowsImage removed for copyright reasons.But where is the sun?And let’s not forgetStereopsis,Vergence,andMotion