Facts and Myths About image Patterns
|
|
|
- Ambrose Moody
- 5 years ago
- Views:
Transcription
1 CAR-TR-968 CS-TR-4251 CorneliaFermuller1,HenrikMalm2andYiannisAloimonos1 StatisticsExplainsGeometricalOpticalIllusions 1CenterforAutomationResearchIIS CollegePark,MD UniversityofMaryland May2001 2MathematicalImagingGroup(MIG) DepartmentofMathematics(LTH) LundInstituteofTechnology/ S-22100Lund,Sweden LundUniversity P.O.Box118 estimatedfromtheinputdata.however,noiseintheimageintensityanditsderivativescauses Itisproposedinthispaperthatmanygeometricalopticalillusions,aswellasillusionary problemsintheestimationofthefeatures;inparticular,itcausesbias.asaresult,thelocations intersectionsoflines,orlocalimagemovementmustbederived.thatis,thesefeaturesmustbe patternsduetomotionsignalsinlinedrawings,areduetothestatisticsofvisualcomputations. Theinterpretationofimagepatternsisprecededbyastepwhereimagefeaturessuchaslines, Abstract Thisworkrevealsageneraluncertaintyprinciplegoverningtheworkingsofvisionsystems.A offeaturesareperceivederroneouslyandtheappearanceofthepatternsisaltered.thebias occurswithanyvisualprocessingoflinefeatures;underaverageconditionsitisnotlargeenough tobenoticeable,butillusionarypatternsaresuchthatthebiasishighlypronounced.ingeneral, thisbiascannotbeavoided,andanyvisionsystem,biologicalorarticial,mustcopewithit. consequenceofthisprincipleisthepredictionofalargeclassofgeometricopticalillusions. acknowledged,asisthehelpofsaralarsoninpreparingthispaper. ThesupportofthisresearchbytheNationalScienceFoundationundergrantIIS isgratefully
2 thestreets.everybookstorehasasectiononthetopic,andmanycuriosityandsouvenirstores carryparaphernaliawithdecorativeillusionarypatterns,suchasthedrawingsofm.c.escher Opticalillusionsareasourceoffascinationtomanypeople.Almosteveryoneisfamiliarwith 1Introduction someillusions,asinrecentyearstheyhaveinltratedthepopularsciencesaswellastheartof ancientgreecedevelopedaverygoodunderstandingofthedistortioneectsinstructuresof numberofdierentexplanationsoeredisevenlarger. probablygettheimpressionthattherearehundredsofseeminglyunrelatedillusions,andthe ationofillustrativeartandexplanatorymaterial.butifonetriestondexplanationshewill orgureswithdoubleinterpretations. largeextentandtheyappliedcompensatorymeasures.asverticalcolumnswouldappearto projects,interactivedemonstrations,lotsofartwork,andevencompaniesdevotedtothecre- Someopticalillusionshavebeenknownsinceantiquity.Forexample,thearchitectsof IfonesearchestheInternet,onewillndmanycollectionsofopticalillusions,manyschool beconvextowardtheground,theyconstructedthemtobeconcave;astallcolumnsarelikely leanoutwardatthetop,theyinclinedtheminward;aslonghorizontalbeamswouldappearto toappearsomewhatshrunkeninthemiddle,theymadethemslightlyswollen.anothereect, appearssmalleratthezeniththanatthehorizon.ptolemyprovidedanexplanationforthis documentedveryearlyinhistory,isthemoonillusion,thatis,thephenomenonthatthemoon canbefoundinhelmholtz'streatiseonphysiologicaloptics[34].sincethen,illusionshave eectthatisstillconsideredtoday. sionsbecamethesubjectofsystematicstudy.alargepartofthendingsofthelastcentury beenstudiedmostlyintheeldofpsychology,andovertheyearstheyhavereceivedvarying amountsofattention.inthelastdecades,withgrowinginterestinthestudyofthemindand eldssuchasneurology,physiology,philosophy,andthecomputationalsciences,inparticular perception,illusionshavealsobeenlookedatoccasionallybyscientistsinothermind-related bythoseinterestedinthemodelingofneuralnetworks. Inthenineteenthcenturyscientistsstartedtobeinterestedinperception,andopticalillu- revealsomethingabouthumanlimitationsandbytheirnatureareobscureandthusfascinating. Butthishasnotbeenthesolereasonforscientists'interest.Fortheoristsofperceptiontheyhave beenusedastestinstrumentsfortheory;thiseortoriginatedfromthefoundersofthegestalt school.astrategyinndingouthowcorrectperceptionoperatesistoobservesituationsin whichmisperceptionoccurs.bycarefullyalteringthestimuliandtestingthechangesinvisual performance,psychologistshavetriedtogaininsightintotheprinciplesofperception. Whatisitthathascausedthislong-standinginterestinopticalillusions?Forone,they inuencesonillusionarypatterns thatis,theattitudeoftheobserver,orientation,thickness, numberoflines,etc.,whichcausethemaximumorminimumillusionaryeectinapattern. animals.usingthisknowledgeitispossibletoinventvariationsonillusionarygures.todo oftheobserver'sagehasbeeninvestigated,andafewstudieshaveevenbeenconductedwith thisthereisnoneedtounderstandinanydeepwayhowanillusionworks. Forsomeofthebestknownillusionscross-culturalstudieshavebeenperformed,theinuence Fromalargenumberofexperimentalstudiesalothasbeenlearnedabouttheparametric areanindicatoroftheforlornnessofthehopeofndingageneraltheorywhichwouldaccount edgethatanyexistingtheorysatisfactorilyexplainsopticalillusions,orthattherearemajor principlesunderlyingclassesofillusions.mostwouldagreewithrobinson[29],whowritesin hisintroductiontogeometricalopticalillusionsthatthelargenumberofillusionsinthisclass Atthistime,however,hardlyanyoneinvolvedinthestudyofperceptionwouldacknowl- 1
3 wewillreducetheobservationstoafewprinciples,expressedinonegeneraltheorywhichis capableofpredictingthefailuresaswellasthesuccessesoftheperceptualsystem. forallofthem.butthestudyofperceptionisstillataveryearlystage.forcomparison,the universallaw.thuswehavegoodreasontobelievethatasmoreknowledgebecomesavailable, manyphysicallawsexplainingtheseobservations,andmodernphysicsisnowinsearchofthe muchmorematureeldofphysicsstartedoutinasimilarway,withmanyobservationsand termisatranslationofthegermangeometrisch-optischetauschungenandhasbeenused foranyillusionseeninlinedrawings.itwascoinedbyoppel[26]inapaperaboutthe overestimationofaninterruptedascomparedwithanuninterruptedextent,latercalledthe largeextentnexttoasmalloneandtheunderestimationofasmallextentnexttoalargeone, Oppel-Kundtillusion[19]. andzollnerillusions,thefraserspiral,andthecontrasteect(thatis,theoverestimationofa Thebestknownandmoststudiedofallillusionsarethegeometricalopticalillusions.The theoverestimationinbrightnessofabrightentitynexttoadarkone,etc.).thenumberof bequitediverse. illusionarypatternsthatfallinthisclassisverylarge,andtheunderlyingphenomenaseemto SomeotherveryfamousillusionsinthisclassincludetheMuller-Lyer[23],Poggendor, oflines,orlocalimagemotionmustbederived,thatis,theymustbeestimatedfromtheinput sionillusionarypatternsduetomotionsignalsinlinedrawings,areduetothestatisticsofvisual computations.wheninterpretingapattern,featuresintheimagesuchaslines,intersections data.noiseintheinputcausestheestimationofthefeaturestobebiased.asaresult,the enoughtobenoticeable,butillusionarypatternsaresuchthatthebiasisstronglypronounced. locationsoffeaturesareperceivederroneouslyandtheappearanceofthepatternisaltered.the biasoccurswithanyvisualprocessingoflinefeatures;underaverageconditionsitisnotlarge Theclaimofthispaperisthatmanyofthegeometricalopticalillusions,andalsobyexten- astaticpatternweperformeyemovementsandgatheraseriesofimages(eitherbymoving oneyemovements[4]thedatafromanumberofretinalimagesiscombinedintooneimage theeyefreelyoverthepatternorbyxatingatsomepointonit).accordingtoexperts representation.helmholtzreferstotherepresentationoftheretinalimageastheeldofxation (Blickfeld)andtotherepresentationthatmoveswiththeeyeasthevisualglobe(Sehfeld).More recentlythetermsretinotopicframeandstablefeatureframehavebeenused[7]. Insomewhatmoredetail,oureyesreceiveasinputasequenceofimages.Evenifweview sequencewhichserveasinputtofurtherestimationprocesses.forthemodelingconducted hereweconsiderthreekindsofimagemeasurements:thegrayvaluesofimagepoints;small imagemeasurementscanonlybederivedwithinarangeofaccuracy.inotherwords,there edgeelements(edgels);andimagemotionperpendiculartolocaledges(normalow).these rstinterpretationprocessesconsistofestimatinglocaledgesusingasinputalltheavailable orientationsofedgeelements,andinthedirectionsandlengthsofnormalowvectors.the isnoiseinthepositionsandgrayvaluemagnitudesofimagepoints,inthepositionsand Theearlyvisualprocessingapparatusextractslocalimagemeasurementsfromtheimage imagepoints,orestimatingtheintersectionsoflinesusingalltheedgels,orcomputinglocal 2Dimagemotionusingallthenormalowmeasurementsinasmallregion.Theseestimation processesarebiased.thustheperceivedpositionsofedgelsareshifted,theirdirectionsare tilted,andtheirmovementsareestimatedwrongly. risetotiltedanddisplacedstraightlinesanddistortedcurvesasperceivedinmanyillusionary patterns.inthecaseofmotion,thelocalimagemeasurementsarecombinedinsegmentation pretationprocesses.longstraightlinesorgeneralcurvesarettedtotheedgelsandthisgives Thelocaledgelandimagemotionestimatesserveasinputtothenexthigherlevelinter- 2
4 and3dmotionestimationprocessestointerpretthespatiotemporalcontentofthescene. visualsystem,butrathergeneralcomputationsthatinsomeformmustbeconductedinvisual interpretationprocesses.wechoosetoemploythreedierentmodelsforthreekindsofdata wewillalsoshowthatthebiasexistsforother,moreelaborateestimationprocesses,andthat ingeneralitcannotbeavoided.wedonotattempttomodelthespecicsofthehuman temporalderivatives) forthepurposeofclearandsimpleexplanation.foreachmodelwe discussanumberofillusionsthatarebestexplainedbyit.itshouldbeemphasizedthatthe grayvalues,edgels(spatialderivativesataspatiallocation),andnormalow(spatialand Inthispaperwewilldemonstrateandanalyzethebiasinlinearestimationprocesses,but thedierentiationprocess[25],buttherearemanysourcesofnoisethataectthevisualdata, Lettheirradiancesignalcomingfromthesceneparameterizedbyimageposition(x;y)and dataandnoisyspatiotemporaldatacanbederivedfromnoisygraylevelvaluesbymodeling andthedierentiationprocesscertainlyisoneofthem. 2Erroringrayvalues visionsystem,inalmostanyinterpretationprocess,usesallthreekindsofdata.thenoisyedgel graylevelvalue,andthereisnoiseinspatiotemporallocation.wemodelthesenoisesources asfollows: theidealsignal.ingeneral,twokindsofnoisesourcesmustbeconsidered:thereisnoisein tributednoisetermi(x;y;t)withzeromeanandcovariancematrixg.thisrepresentsthe time(t)bei(x;y;t).theimagereceivedontheretinacanbethoughtofasanoisyversionof bution.theexpectedvalueoftheintensitythusamountstotheconvolutionofthesignalwith erroringraylevelvalue.wemodelitasagaussian,i.e.i(x;y;t)n(0;g).asamodelfor thepositionalerror,weconsiderthevalueatapointashavingagaussianprobabilitydistri- aspatiotemporalgaussiankernel,thatis, Ateveryimagepoint(x;y;t)weconsideranadditive,independentlyandidenticallydis- withx=(x;y;t)and g(x;p)= E(I(x)+N(0;g)g(x;p)) standarddeviationinthegraylevelnoise.theexpectedvalueoftheintensityisthen y-directionstobethesame,withpthestandarddeviationinthepositionalnoiseandgthe forsimplicity,throughoutthepaperweassumethespatialnoisecomponentsinthex-and Inthemodelingofstaticsignalsonlythespatialvariablesneedtobeconsidered.Furthermore, (2)3=2jpj1=2e?12xt?1 1 E(I(x;y)+N(0;g)g(x;y;p))=E(I(x;y)g(x;y;p))+E(N(0;g)g(x;y;p))(1) px oftheintensitycanbewrittenasthesumoftwoterms.thesecondterm,theexpectedvalue second-orderderivatives(thelaplacian,r2)[22].ascanbeseenfrom(1),theexpectedvalue ofn(0;g)g(x;y;p)anditsderivativesareindependentofposition(theyarezero).this meansthatnoiseinthegrayleveldoesnotaecttheexpectedlocationsofedgesandforthe Edgesareobtainedastheextremaoftherst-orderderivatives[3]orasthezerocrossingsof 3
5 purposeofthisanalysisitcanbeignored.1thusweareinterestedinthelocations(x;y)where Beforeanalyzingedgesundersmoothing,letuslookatthesourcesofthenoise.Oneinterpretationisthatthereisuncertaintyinthepositionsinceasequenceofimagesiscombinedinto hasalocalmaximum,orwherer2g(x;y;p)i(x;y)=0 @yi(x;y))2]1=2 onecoordinatesystemanderrorsoccurinthegeometriccompensationforlocation.another withagaussianwithastandarddeviationthatisthesumofthestandarddeviationsofthe obtainedisaprocessedversionoftheidealsignal,whichevenforonlyoneimagetakenatone [22,30].SincerepeatedlteringwithdierentGaussiankernelsisequivalenttoconvolution predictsthatedgesintheearlyvisualsystemarecomputedusingadierenceofgaussian lter,thatis,aband-passlterwhichisapproximatedwellbythelaplacianofagaussian instantintimecanbemodeledbygaussiansmoothing.furthermore,tocomputethederivativessomelow-passltering(thatis,smoothing)isnecessary.forexample,acurrenttheory interpretationisthattheeyecanbeviewedasade-focusedimagingsystemandtheimage thatwecanemploystraightforwardlyinourstudy. scalespacetheory,andalargenumberoftheoreticalresultsonthistopichavebeenderived ofedgepointsinthedirectionnormaltoedges,theso-calleddriftvelocity. kernels,anycombinationoftheabovenoisesourcesisalsowellapproximatedbyourmodel. paralleltothespatialgradientdirectionatp0andtheu-axisperpendiculartoit.ifedgesare ofthesmoothingparameter.lindebergin[20]derivesformulaefortheinstantaneousvelocity Weareinterestedinthepositionsofedges,orthechangeinpositionsofedgeswithvariation ConsiderateveryedgepointP0alocalorthonormalcoordinatesystem(u;v)withthev-axis Gaussiansmoothingofstaticimageshasbeenintensivelystudiedintheliteratureonlinear scaleparameter)amountsto 2(r2Iu)2+(r2Iv)2r2Iu;r2Iv r2?r2i Thedriftvelocityrepresentsthetendencyofthemovementofedgesinscalespace.Toobtain esto Forastraightedge,whereallthedirectionalderivativesintheu-directionarezero,itsimpli- Ivvv(0;1) (3) (2) thedriftvelocity.ifthescaleintervalissmallthedriftvelocityinthesmoothedimageprovides, smoothedimage,wouldrequiretracingtheedgepointoverthescaleintervalandintegrating inmanycases,asucientapproximationtothetotaldrift,andthisiswhatwewillshowin laterillustrations. tobeconsidered:edgesbetweenadarkandabrightregiondonotchangelocationunderscale thetotaldriftofanedgepointp0intheoriginalimagetothecorrespondingedgepointinthe intensitypropagatesintolargeruncertaintyintheestimationofthederivativesoftheintensity.thismighthave someeectonhumanperception,anissuethatcouldbeinvestigated. 1Noiseinthegraylevelaectsthesecondmoment,thatis,thevariance.Uncertaintyintheestimationof ThescalespacebehaviorofstraightedgesisillustratedinFigure1.Therearethreecases 4
6 twoedgestowardeachother.theseobservationssucetoexplainanumberofillusions.the spacesmoothing(figure1a).thetwoedgesattheboundariesofabrightline,orbar,ina underlyingmathematicsisexplainedinfigure2. ofsmoothingonalineofmediumbrightnessnexttoabrightandadarkregionistomovethe parameterislargeenoughthatthewholebaraectstheedges(figure1b).finally,theeect darkregion(or,equivalently,adarklineinabrightregion)driftapart,assumingthesmoothing Figure1:Aschematicdescriptionofthebehaviorofedgemovementinscalespace:(a)no movement,(b)driftingapart,(c)gettingcloser. illusionarypatternsareshownwhichconsistofregulargridsorcheckerboardpatternswithlittle squaressuperimposedonthem.theeectofthesquaresistochangetheappearanceofthe AttheWebsiteoftheAkitaUniversityDepartmentofPsychology[1],anumberofnew (b) (c) blacksquaressuperimposed,andmanyofthesesquaresalignedwiththegrid.figure3bshows asmallpartoftheguremagnied.theblacksquaresinthecenterofthegridallhave straightlinestoconcaveandconvexcurves. beenremovedforclarity,astheydonotnotablyaecttheillusionaryperception.figure3c Figure3dshowsthesmoothedimagewhichresultsfromlteringwithaGaussianwithstandard showstheresultsofedgedetectionontherawimageusingthelaplacianofagaussian(log). deviation5/4timesthewidthofthebarsandfigure3eshowstheresultofedgedetectionon thesmoothedimageusingalog.(thisisclearlythesameasperformingedgedetectiononthe Figure3ashowsoneofthegures,ablacksquaregridonawhitebackgroundwithsmall originallocationswherethegridboundslittlesquaresandareshiftedinsidethewhiteblocks rawimagewithalogoflargerstandarddeviation.) atallotherlocations,andasaresulttheedgesappeartobecurved(figure3e).ascan todriftthetwoedgesapart.atthelocations,however,whereasquareisalignedwiththe grid,thereisonlyoneedge,andthisedgedoesn'tchangelocationunderscale-spacesmoothing. Theneteectofsmoothingisthatedgesofgridlinesarenolongerstraight;theystayatthe beveried,thecurvatureobtainedfromscale-spacesmoothingisqualitativelysimilartothe curvatureperceivedinthegure. Thegridconsistsoflines(orbars),andtheeectofscalespacesmoothingonthebarsis 5
7 Figure2:Therstrowshowsgraylevelguresof(a)adarkregionnexttoabrightregion,(b)a brightlineinadarkregion,and(c)agraylinenexttoadarkandabrightregion.superimposed (b) =7).Thethirdrowshowstherstderivatives,andthefourthrowthesecondderivatives onthegraylevelguresaretheedgeswhichhavebeencomputedasthezerocrossingsofthe versionsofthegraylevelgures(thesolidlinecorrespondsto=3andthedashedlineto toastandarddeviation()ofthreeandthedarkerlinestoaofsevenpixels).thesecond LaplacianofaGaussianfortwodierentwidthsoftheGaussian(thebrighterlinescorrespond rowshowsthefunctionsresultingfromcrosssections(onerow)ofthetwodierentlysmoothed 0.01 ofthesefunctions.themaximaoftherstderivativesandthezerocrossingsofthesecond derivativesaremarkedbyverticallines.fromtheirchangeinpositiontheobservationsabout themovementofedgesinscalespaceasshowninfigure1canbededuced
8 (a) (b) (c) Figure3:(a)Illusionarypattern:\spring."(b)Smallpartoftheguretowhich(c)edge detection,(d)gaussiansmoothing,and(e)smoothingandedgedetectionhavebeenapplied. 7 (e)
9 checkerboardtile).thisexplainstheeect:theedgeswhicharetheboundariesofthecreated barsdriftapartunderscale-spacesmoothingandtheotheredges betweentheblackandwhite tilesofthecheckerboard stayinplace.theresultisthattheedgesnearthelocationsofthe boardwithlittlewhitesquaressuperimposedincornersoftheblacktilesclosetotheedges. square)nexttoablackbar(fromablackcheckerboardtile)nexttoawhitearea(fromawhite whitesquaresarebumpedoutwardtowardthewhitecheckerboardtiles.thisisillustrated Inthispattern,nexttothewhitesquaresshortbarsarecreated awhitearea(fromalittle Figure4ashowsanotheroneofthepatternsfromtheWebsite:ablackandwhitechecker- from(2).againtheeectobtainedfromsmoothingisthesameastheoneperceivedwiththis infigure4bwhichshowsthecombinedeectofsmoothingandedgedetectionforapartof linesseparatingtherows.themortarlinesshouldbemid-wayinluminancebetweentheblack illusionarypattern. Aninteractivedemonstrationofthisillusioncanbefoundat[17].Itconsistsofablackand andwhitesquares;thentheillusionaryeectofatiltinthemortarlines,withalternatemortar whitecheckerboardpatternwithalternaterowsshiftedonehalf-cycleandwiththinmortar thepattern.figures4canddzoominonthedriftvelocityinthesmoothedimageasderived linestiltedinoppositedirections,isstrongest.thecafewallgureiscloselyrelatedtothe Munsterberggure[24],whoseonlydierenceisthatthemortarlinesarethesameluminance aseitherthelightordarktiles;inthisgurethetiltingisperceivedtobemuchweaker. Anotherillusionarypatterninthiscategoryisthe\cafewall"illusionshowninFigure5a. brighttilethetwoedgesmovetowardeachotherunderscalespacesmoothing,andforthinlines ittakesarelativelysmallamountofsmoothingforthetwoedgestomergeintoone.where themortarlineisbetweentwobrightregionsorwhereitisbetweentwodarkregionstheedges moveawayfromeachother.theresultsofsmoothingandedgedetectionareshowninfigure 5cforasmallpartofthepatternshowninFigure5b.Fromthisgureitcanbeseenthatthe edgeelementswhichformboundariesofthetiles(thatis,edgeelementsthatformthelower boundaryofawhitetileandtheupperboundaryofablacktile,andedgeelementsthatform Inthecafewallgure,atthelocationswhereamortarlinebordersbothadarktileanda samesignofslopeasperceived.figures5dandeshowthedriftvelocityinlargemagnication thelowerboundaryofablacktileandtheupperboundaryofawhitetile)aretiltedwiththe tile,therewillbeanedgecloserinpositiontotheblacktilethanwiththegraymortarline. themunsterbergillusioniseasilyunderstood.if,forexample,themortarlineischangedfrom graytowhitetherewillbenoedgesbetweenwhitetiles,andattheborderofawhiteandblack whichillustratesthetendencyofthemovementoftheedges.theweakerillusionaryeectin showsasmallpartofthepatternandfigure6cshowstheedgesdetected.asthelattergure shows,theinsertedsquarespartlycompensateforthedriftinginoppositedirectionsofedges whiteandblacksquaresputinthecornersofthetilesremovetheillusionaryeect.figure6b Theresultisfeweredgeelementswithlesstiltandthusaweakerillusion. andasaresultastraightlinewithouttiltisseen. alongthemortarlineseparatingtilesofthesamegraylevel.asaresultslightlywavyedgels areobtained;butthe\waviness"istooweaktobeperceived(lowamplitude,highfrequency) eectbyintroducingadditionalelements.thishasbeenpursuedinfigure6a;theadditional Ifbiasisindeedthemaincauseoftheillusionthenweshouldbeabletocounteractthe explainsthetiltingoftheseelements.thesecondstageconsistsoftheintegrationofthese localelementsintolongerlines.ourhypothesisisthatthisintegrationiscomputationallyan isduetotwoprocessingstages.intherststagelocaledgeelementsarecomputedandbias approximationofthelongerlinesusingasinputthepositionsandorientationsoftheshortline Afullaccountoftheperceptionoftiltedlinesrequiresadditionalexplanation.Theillusion 8
10 (a) (b) Figure4:(a)Illusionarypattern:\waves."(b)Theresultofsmoothingandedgedetection onapartofthepattern.(c)and(d)thedriftvelocityatedgesinthesmoothedimage logarithmicallyscaledforpartsofthepattern. (d) 9
11 (a) (b) (c) Figure5:(a)Cafewallillusion.(b)Smallpartofthegure.(c)Resultofsmoothingandedge detection.(d)and(e)zoom-insonthedriftvelocity. 10 (e)
12 (a) inthelteredimage,whichcounteractstheillusionaryeect. Figure6:Modiedcafewallpattern.Theadditionalblackandwhitesquareschangetheedges (b) (c) 11
13 spaceandcouldberealizedinamultiresolutionarchitecture.ateveryresolutiontheaverage ofthedirectionsofneighboringelementsiscomputed,andallcomputationsarelocal.ifthe elements.thiscouldbeeasilyimplemented.itbasicallyamountstoasmoothingindirection becomputedashavingatilt. approximationiscarriedoutonthebasisoftheorientationsoflocaledgeelementsthelinewill informationabouttheedgeelementswhichwasderivedearlierandgobacktoimagepoint asinputonlythepositionsofimagepointsontheedgels;suchanapproximationwouldnot information.wedoubtthatevolutionwouldproduceawastefulapproachofthiskind. leadtoatiltintheestimatedlines.however,insuchacasethesystemwouldthrowoutthe Theonlyotherwayofcarryingouttheintegrationwouldbetoapproximatethelines,using twistedcordgivestheperceptionofbeingspiral-shaped,ratherthanasetofcircles.the appearsinbothblackandwhite.inthefraserillusioninfigure8theshortlinesarealltilted principleofthepatterniseasilyexplainedusingthefraserillusioninfigure8.figure8c showsthebasicunitofallfraser'sgures,ashortlinewithatriangleateitherend.this whichtogetherformsomethingratherlikeatwistedcord,onacheckerboardbackground.the Fraser'stwistedcordillusion,ortheFraserspiral,andrelatedFrasergures. TheFraserspiralpattern(Figure7)consistsofcirclesmadeofblackandwhiteelements Ouraccountoftheintegrationprocessalsoexplainsoneofthemostforcefulofallillusions, lines(figure8)orspirals(figure7).wecanalsoaccountfortheincreaseintheperceptionof lines.anapproximationprocessthatusestheseedgelstotextendedcurveswillderivetilted withthesameslopeandatiltedlineisperceived;inthefraserspiralthelineelementsare lineelementsandtheintersectionoftheselineelementswiththetiltedlineelementsofthe thetiltwithinclusionofthebackground.inthiscasetheadditionaltrianglescreateadditional cordonagraybackground(figure8b)butitisslightlyweakerwithoutthepatternofthe backgroundincluded. sectionsofspiralsandspiralsareperceived.theillusionaryeectisalsoobtainedforatwisted gureisbiased,andthisincreasesthetiltofthelineelements. themodelintroducedinthissectionthattheintersectionpointoftwolineswhichintersectat TheperceptualeectatintersectinglinesisillustratedinFigure9.Itcanbeshownwith TheedgeswhicharecomputedinFraser'sguresareatthebordersoftheblackandwhite 3Errorsinlineelements anacuteangleisdisplaced.theeectisobtainedbysmoothingtheimageandthendetecting ofintersectinglinesisthetopicofthenextsection. edgesusingnon-maximumsuppression(seefigure10).amoredetailedanalysisofthebehavior slightlydierentnoisemodel,withthenoisebeingdeneddirectlyontheedgeelements.we areadecisivefactorintheillusion.toanalyzetheselinedrawings,weadoptinthissectiona Thereisalargegroupofillusionsinwhichlinesintersectingatangles,particularlyacuteangles, thecellsanalyzingorientation[15]. edgeelements,parameterizedbytheimagegradient(avectorinthedirectionnormaltothe Anotherphysiologicallyplausibleexplanationisthatsuchnoiseisduetoquantizationerrorsof thatonlythenoiseinthedirectionhasaninuenceonthepositionsoftheintersectionpoints givethesameeects.noiseingraylevelvaluesresultsinnoiseintheestimatededgeelements. assumeadditivenoiseinthepositionsanddirectionsofedgeelements.letusnotebeforehand edge)(ix;iy)andthepositionofthecenteroftheedgeelement(x0;y0). oflines,andthusamodelthatdoesn'tconsideranynoiseinthepositionsofedgeelementswill Wenextanalyzetheestimatedpositionofanintersectionofstraightlines.Theinputsare 12
14 Figure7:Fraser'sspiral[9]. ateitherend. Figure8:From[29].(a)Frasergurewithstraighttwistedcords.(b)Theillusionisweaker withagraybackground.(c)thebasicelementsofthegurearetiltedelementswithatriangle (b) 13
15 Figure9:[34]ThenelineasshowninAappearstobebentinthevicinityofthebroader blackline,asindicatedinexaggerationinb. Figure10:(a)Alineintersectingabaratanangleoffteendegrees.(b)Theimagehasbeen smoothedandthemaximaofthegraylevelfunctionhavebeendetectedandmarkedwithstars. (c)magnicationofintersectionarea. (b) (c) 14
16 rameters.inthesequelunprimedlettersareusedtodenoteestimates,primedletterstodenote actualvalues,and'stodenoteerrors,whereix=i0x+ix,iy=i0y+iy,x0=x0+x0and y0=y0+y0. Consideradditive,independentlyidenticallydistributed(i.i.d.)zero-meannoiseinthepa- Thisequationisapproximatedbythemeasurements.Letnbethenumberofmeasurements. Eachmeasurementiprovidesoneequation Foreverypoint(x;y)onthelinesthefollowingequationholds: andweobtainasystemofequationswhicharerepresentedinmatrixformas Ixix+Iyiy=Ixix0i+Iyiy0i I0xx+I0yy=I0xx0+I0yy0 (4) HereIsisthen-by-2matrixwhichincorporatesthedataintheIxiandIyi,and~Cisthe n-dimensionalvectorwithcomponentsixix0i+iyiy0i.thevector~xdenotestheintersection Is~x=~C (5) leastsquare(ls)estimationisgivenby~x=(itsis)?1its~c pointwhosecomponentsarexandy.thesolutiontotheintersectionpointusingstandard ydirectionsisassumedtobethesame,letitbe2s,andalsotheexpectedvaluesofhigher- twoothernoisytermsandthusthestatisticsaresomewhatdierent. themeasurementmatrixaisbiased.thestatisticsoftheestimationhavebeenstudiedforthe Itiswellknown[10]thattheLSsolutiontoalinearsystemoftheformA~x=~bwitherrorsin caseofi.i.d.noiseintheparametersofaand~b.inourcase~bistheproductoftermsinaand Tosimplifytheanalysis,thevarianceofthenoiseinthespatialderivativesinthexand (6) ~xisfoundbydeveloping(6)intoasecond-ordertaylorexpansionatzeronoise.itconverges (thansecond)ordertermsareassumedtobenegligible.intheappendixtheexpectedvalueof ~x0istheactualintersectionpointand~x0="1npni=1x0i inprobabilityto where M0=Is0tIs0="Pni=1I02xiPni=1I0xiI0yi n!1e(~x)=~x0+nm0?1(~x0?~x0)2s plim Pni=1I0xiI0yiPni=1I02yi# 1nPni=1y0i#isthemeanofthe~x0i. (7) (~x0?~x0).vector(~x0?~x0)extendsfromtheactualintersectionpointtothemeanpositionof eigenvectorsareorthogonaltoeachother.m0?1hasthesameeigenvectorsasm0andinverse shiftedbyatermwhichisproportionaltotheproductofmatrixm0?1andthedierencevector whichdependsonlyonthespatialgradientdistribution,isarealsymmetricmatrixandthusits theedgeelements.thusitisthemasscenteroftheedgelsthatdeterminesthisvector.m0, eigenvalues.thedirectionoftheeigenvectorcorrespondingtothelargereigenvalueofm0?1is dominatedbythenormaltothemajororientationoftheimagegradientsandthustheproduct Using(7)allowsforaninterpretationofthebias.Theestimatedintersectionpointis ofm0?1withvector(~x0?~x0)ismoststronglyinuencedbythisorientation.thustheeect 15
17 thisgureisapparentlynotthecontinuationofthelowerportionontheright,butistoohigh. interactiveversionsee[16]).theupper-leftportionoftheinterrupted,tiltedstraightlinein AnotherversionofthisillusionisshowninFigure11b.Hereitappearsthatthemiddleportion ofmoregradients.forthecaseoftwointersectinglinesthismeansmoredisplacementofthe ofm0?1ismorebiasinthedirectionoffewerimagegradientsandlessbiasinthedirection AversionofthePoggendorillusionasdescribedbyZollnerisdisplayedinFigure11a(foran oftheinclined(interrupted)lineisnotinthesamedirectionasthetwoouterpatterns,butis intersectionpointinthedirectionperpendiculartothelinewithfeweredgeelements. turnedclockwisewithrespecttothem. ThebestknownillusionsduetointersectinglinesarethePoggendorandZollnerillusions. illusionsastheresultoftheperceptualenlargementofacuteangles.ithasbeenfoundinextensivestudiesthatacuteanglesareoverestimated,andobtuseanglesareslightlyunderestimated Famousvisionscientists[13,34]sawthePoggendoraswellastheZollnerandotherrelated (a) Figure11:Poggendorillusion. (b) (althoughregardingthelattertherehasbeencontroversy).theeectislargerwithshortline segmentsandseemstodiminishwithlongintersectinglines.ourmodelpredictsthatthese phenomenaareduetothebiasintheestimationoftheintersectionpoint. lineismoveddownandtotheright.asaresultthetwolinesegmentsappeartobeshiftedin ismovedupandtotheleft,andtheintersectionpointoftherightverticalwiththelowertilted oppositedirectionsandnottolieonthesamelineanymore. plottheestimatedbiasinthex-andy-directionsasafunctionoftheanglebetweentheline elements.inthisplotthedataisaverticallinewhichmakesananglewithatiltedline,and intheacuteangleandreacheszeroat90degrees.ourmodelpredictsthis.figures12bandc ReferringtoFigure11a,theintersectionpointoftheleftverticalwiththeuppertiltedline thenumberofedgeelementsontheverticallineistwicethatonthetiltedline,asillustrated thediagonallinesinfigure13bareallparallel,buttheylookconvergentordivergent.inthese infigure12a. Fromparametricstudiesitisknownthattheillusionaryeectdecreaseswithanincrease Figure13showstwoversionsoftheZollnerillusion.TheverticalbandsinFigure13aand 16
18 φ Figure12:(a)Thetiltedlineintersectstheverticallineinthemiddleatanangle.Thereare vertical.(c)biasparalleltothevertical. twiceasmanyedgeelementsontheverticalasonthetiltedline.(b)biasperpendiculartothe (c) bias in x bias in y tiltedbars. samedirectionasperceivedbythevisualsystem.figure14showstheestimationofthetilted lineelementsforapatternsuchasinfigure13awith45degreesbetweentheverticalandthe tilted,asillustratedinfigure13c.inasecondcomputationalstep,longlinesarecomputedas anapproximationtothesmalledgeelements,andthisgivesrisetotiltedlinesorbarsinthe linesegmentscausetheedgeelementsalongthelongedgesbetweenintersectionpointstobe patternsthebiasesintheintersectionpointsofthelonglines(oredgesofbands)withtheshort phi phi thelineelementsbetweenintersectionpointstobetilted. Figure13:(a),(b)Zollnerpatterns.(c)Thebiasintheintersectionpointsoftheedgescauses increasinglyacuteanglebetweenthemainlineandtheobliques,whichcanbeexplainedas beforeusingfigures12bandc(or,similarly,figure17).thevaluewherethemaximumoccurs Inexperimentswiththisillusionithasalsobeenfoundthattheeectdecreaseswithan (b) (c) variesamongdierentstudies.itissomewherebetween10and30degrees;belowthat,some 17
19 counteractingeectsseemtotakeplace. Figure14:EstimationofedgesinZollnerpattern.Thelineelementsarefoundbyconnecting twoconsecutiveintersectionpoints,resultingfromtheintersectionofedgesoftwoconsecutive timesmoreelementsonthevertical) acongurationsimilartotheoneinfigure17. consistsofedgeelementsuniformlydistributedontheverticalandonthetiltedline(with1.5 barswiththeedgeoftheverticalbar(oneinanobtuseandoneinanacuteangle).thedata strongestwiththeparallellinesverticalorhorizontal.thezollnerillusion,ontheotherhand, wasfoundtobemaximalwhenthejudgedlinesareat45degrees,asinfigure13b. ofthepoggendorandzollnergures.somestudiesfoundthatthepoggendorillusionis Otherparametricstudieshavebeenconductedontheeectofalteringtheorientations lines,thatis,moreedgeelementsareestimatedintheseandnearbydirections.(inthepatterns zontaldirections:wegenerallyseebetterintheseorientations.anatomicalstudies[15]have tations[21].basedonthesendingsweassumethereismoredataforhorizontalandvertical foundorientation-selectivecellsinthecortex;dierentcellsrespondtodierentangles.ithas alsobeenfoundthatthereismoreactivityforhorizontalandverticalthanforobliqueorien- consideredtheremaybemoreactivitybecauseofmoreeyemovementsalongthesedirections, asthevisualtasksrequirejudgmentsalongthesedirections.) Psychophysicistshavecoinedtheterm\spatialnorms"torefertotheverticalandhori- distributiononthebiasisstrongestinthedirectionoftheeigenvectorcorrespondingtothe inthedirectionsofthespatialnorms.inadirectionwheretherearemoreestimatesthereisa largersignaltonoiseratio,andthisresultsingreateraccuracyinthisdirection. Asimplestatisticalobservationexplainsthehigheraccuracyintheestimationofentities y-directionincreases.thepoggendorillusionisstrongerwhentheparallellinesarevertical verticaltoobliqueelementsincreases,thebiasinthex-directiondecreasesandthebiasinthe largereigenvalueofm0?1andweakestintheorthogonaldirection.changingtheratioof measurements(edgels)alongthedierentlineschangesthebias.figure15illustrates,forthe andy-directionsasafunctionoftheratioofedgeelements.itcanbeseenthatastheratioof caseofaverticallineintersectinganobliquelineatanangleof30degrees,thebiasinthex- Expressedinourformalismof(7),thistakesthefollowingform.Theeectofthegradient andthemainlinesaretilted,asinthiscasethebiasperpendiculartothemainlines(alongthe islarger,andthezollnerillusionisstrongerwhenthesmalllinesarehorizontalandvertical orhorizontal,becauseinthiscasethebiasparalleltothelines(alongthey-axisintheplot) 18
20 x-axisintheplot)islarger. Figure15:(a)Averticallineandatiltedlineintersectingatanangle.(b)Biasperpendicular totheverticalasafunctionoftheratioofedgeelementsontheverticalandtiltedlines.(c) ExamplesaretheOrbisonguresandthepatternsofWundt,Hering,andLuckiesh.Inthese Biasparalleltothevertical. Manyotherwell-knownillusionscanbeexplainedonthebasisofbiasedlineintersection. (b) patternsgeometricalentitiessuchasstraightlines,circles,trianglesorsquaresaresuperimposedondierentlytiltedlines,andtheyappeartobedistorted.thisdistortionresultsfrom theerroneousestimationofthetiltinthelineelementsbetweenintersectionpointsandthe approximatedbystraightlines,andtheratioofelementsbetweenthecircleandthebackground andtwototheouterones.arcsonthecirclebetweentwoconsecutivebackgroundlineswere edges,andwecomputedtheintersectionbetweenanybackgroundedgeandcircleedge(using withastraightlinefourintersectionpoints,twocorrespondingtotheinneredgesofthecircle edgewas2:1(resultinginacongurationsimilartothatoffigure17withthearccorresponding congurationssuchasthoseinfigure15a).thisprovidedforeveryintersectionofthecircle subsequentttingofcurvestotheselineelements. tothevertical).consecutiveintersectionpoints oneoriginatingfromanobtuseandonefrom Figure16illustratestheestimationofthecurveintheLuckieshpattern.Eachlinehastwo theouterlinesegments(onlyininterestingplaces).thisresultedinacurveliketheonewe perceive,withthecirclebeingbulbedoutontheupperandlowerleftandbulbedinonthe anacuteangle wereconnectedwithstraightlinesegments.thenbeziersplineswerettedto upperandlowerright. illusionswillresultintheoverestimationofacuteanglesforshortlinesegments.theunderestimationofobtuseanglescanbeexplainedifweassumeanunequalamountofedgeldataonthe thebiasinthey-directionincreases(withincreasingangle)forobtuseangles,andthisresults elementsontheverticalthanonthetiltedline.thebiasinthex-directionchangessignand Nextletuslookattheerroneousestimationofangles.Thebiasdiscussedfortheabove inasmallunderestimationofobtuseangles. twoarcs.figure17illustratesthebiasforacuteandobtuseanglesforthecaseofmoreedge verticallinesarethesamelength,buttheoneontheleftwithterminalspointingoutward isthemuller-lyerillusion.theoriginalmuller-lyergureisshowninfigure18.thetwo appearsconsiderablylongerthantheoneontherightwithinward-pointingterminals.inthis gure,theestimatedintersectionpointsoftheterminalsandthestraightlineareoutwardin theleftandinwardintherightgure.thiscertainlycouldbeacontributingfactorinthe Theillusionwhichprobablyhasbeenstudiedthemostinthepsychophysicalliterature 19 φ bias in x, phi= Pi/6 3 Weight bias in y, phi =Pi/6 3 Weight
21 (a) (b) Figure16:EstimationofLuckieshpattern:(a)thepattern:acirclesuperimposedonabackgroundofdierentlyarrangedparallellines,(b)ttingofarcstothecircle,(c)magniedupper leftpartofpatternwithttedarcssuperimposed,(d)intersectionpointsandttingofsegments (c) Figure17:(a)Biasperpendiculartothevertical.(b)Biasparalleltotheverticalforlines toouterintersections.phi intersectingatangle(asinfigure15);theratioofverticaltotiltededgeelementsis bias in x, weight = bias in y, weight =3 1.5 phi 2 2.5
22 illusion,buttheeectofthebiasdoesnotseemtobelargeenoughtofullyexplainthisstrong illusion.itislikelythatthereareadditionalcausesforthisillusion,whichareduetohigherlevelinterpretationprocesses.theillusionalsooccurswithotherterminals;itseemstohappen whenaspatialextentisboundedbycontoursthatareeitheroutwardlyconvexorconcave, eithercurvedorangular.forotherterminalsonewouldhavetotcurvestotheedgelsalong theterminalsandcomputetheirintersectionswiththestraightlines. 4Errorsinmotion Whenprocessingimagesequencessomerepresentationofimagemotionmustbederivedas arststage.itisbelievedthatthehumanvisualsystemcomputestwo-dimensionalimage measurementswhichcorrespondtovelocitymeasurementsofimagepatterns,calledoptical Figure18:Muller-Lyerillusion. perpendiculartolinearfeaturesarecomputedfromlocalimagemeasurements.inthecomputationalliteraturethisone-dimensionalvelocitycomponentisreferredtoas\normalow"and theambiguityinthevelocitycomponentparalleltotheedgeisreferredtoasthe\aperture ow.theresultingeldofmeasurements,theopticaloweld,representsanapproximation problem."inasecondstagetheopticalowisestimatedbycombining,inasmallregionof totheprojectionoftheeldofmotionvectorsof3dscenepointsontheimage. isbiased. theimage,normalowmeasurementsfromfeaturesindierentdirections,butthisestimate Opticalowisderivedinatwo-stageprocess.Inarststagethevelocitycomponents velocityofanimagepointinthex-andy-directionsby~u=(u;v),thefollowingconstraint isobtained: derivativesoftheimagegrayleveli(x;y;t)byix;iy,thetemporalderivativebyit,andthe isthatimagegrayleveldoesnotchangeoverasmalltimeinterval.denotingthespatial Weconsideragradient-basedapproachtoderivingthenormalow.Thebasicassumption thedirectionofthegradient[14].weassumetheopticalowtobeconstantwithinaregion. Eachofthenmeasurementsintheregionprovidesanequationoftheform(8)andthuswe Thisequation,calledtheopticalowconstraintequation,denesthecomponentoftheowin Ixu+Iyv+It=0 21
23 obtaintheover-determinedsystemofequations whereisdenotes,asbefore,thematrixofspatialgradients(ixi;iyi),~itthevectoroftemporal derivatives,and~u=(u;v)theopticalow.theleast-squaressolutionto(9)isgivenby Asanoisemodelweconsiderzero-meani.i.d.noiseinthespatialandtemporalderivatives.As ~u=?(itsis)?1its~it: Is~u+~It=0; (9) thetwodirectionsandweassumethathigherthansecond-ordernoisetermscanbeignored. expectedvalueoftheow,usingasecond-ordertaylorexpansion,isderivedintheappendix; intheprevioussection,weassumeequalvariance2sforthenoiseinthespatialderivativesin itconvergesinprobabilitytoplim Thestatisticsof(10)arewellunderstood,astheseareclassicallinearequations.The ThebiasisproportionaltothenegativeproductofM0?1withtheactualow.Inthecase where,asbefore,theactualvaluesaredenotedbyprimes. ofauniformdistributionofimagegradientsintheregionwheretheowiscomputed,m0 thecomputedopticalow;thereisanunderestimation.inaregionwherethereisaunique (andthusm0?1)isamultipleoftheidentitymatrix,leadingtoabiassolelyinthelengthof Equation(11)isverysimilarto(7)andtheinterpretationgiventhereapplieshereaswell. n!1e(~u)=~u0?n2sm0?1~u0; gradientvector,m0willbeofrank1;thisistheapertureproblem.inthegeneralcase,the eigenvectorwiththelargereigenvalueisinthedirectionoffewermeasurements;thusthere ismoreunderestimationinthisdirectionandlessunderestimationinthedirectionofmore measurements.theestimatedowthereforeisbiaseddownwardinsizeandbiasedtowardthe majororientationofthegradients. thebiascanbeunderstoodrathereasily.theeigenvectorsofm0?1areinthedirectionsof entationsurroundinganinnerring.smallretinalmotions,orslightmovementsofthepaper, manynormalowmeasurementsinonedirectionastheother.forsuchagradientdistribution oflengthtowidthisfour,leadingtoagradientdistributioninasmallregionwithfourtimesas causeasegmentationoftheinsetpattern,andmotionoftheinsetrelativetothesurround. oftworectangularcheckerboardpatternsorientedinorthogonaldirections abackgroundori- Thetilesusedtomakeupthepatternarelongerthantheyarewide;inFigure19theratio Figure19showsavariantofapatterncreatedbyHajimeOuchi[27].Thepatternconsists Figure20b,whichhasitsminimumat0anditsmaximumat=2(thatis,when~u0isaligned withthemajorgradientdirection).theerrorinangleisgreatestfor=4(thatis,when~u0is opticowandthedominantgradientdirection.theplotsarebasedontheexactsecond-order Figure20showstherelationshipbetweentheopticowbiasandtheanglebetweentheactual thetwogradientmeasurements,withthelargereigenvaluecorrespondingtofewergradients. nearlyeliminatedwhenitisalignedwiththeow(thatis,intheouchipattern,whenthelong edgeoftheblockisperpendiculartothemotion).thebiasforanglesbetween=2and exactlybetweenthetwoeigenvectorsofm0?1)anditis0for0and=2(figure20c).overall, Taylorexpansion. thismeansthebiasislargestwhenthemajorgradientdirectionisnormaltotheowandis As~u0=(0;1),thenoisetermin(11)leadstoabiasinlength,asshownbythecurvein 22
24 Figure19:ApatternsimilartotheonebyOuchi. 23
25 fewer measurements (a) y u 0 θ more measurements x betweentheexpectedowandtheactualow.theerrorhass=0:15. andoflength1.(b)expectederrorinlength.(c)expectederrorinangle,measuredinradians, Figure20:(a)16measurementsareinthedirectionmakinganglewiththepositivexaxis and4measurementsareinthedirection+=2.theopticalowisalongthepositiveyaxis (c) theta theta 24
26 errorintheangle. gradientdirectiondierintheinsetandthesurround,sotheregionalvelocityestimatesare isobtainedfromtheaboveplotsbyreectingthecurvesin=2andchangingthesignofthe forallanglesthedierencebetweentheerrorvectorintheinsetandtheerrorvectorinthe biasedindierentways.when,insteadoffreelyviewingthepatternoffigure19,thepage ismovedindierentdirections,weobservethattheillusionarymotionoftheinsetismostly surroundingareaprojectedonthedominantgradientdirectionoftheinsetisinthisdirection. anglewiththemotionofthepaperislessthan90.usingfigure20,itcanbeveriedthat aslidingmotionorthogonaltothelongeredgesoftherectangleandinthedirectionwhose IntheOuchipattern,therelativeanglesbetweentherealmotionandthepredominant aperceivedmotiontotheright.ifthemotionofthepaperisupward,thedierencevectoris downward;itsprojectiononthemajorgradientdirectionoftheinsetiszero,andthushardly anyillusionarymotionisperceived.figure21shows,forasetoftruemotions,thebiasesin Forexample,whenthemotionisalongtherstmeridian(totherightandup),theerrorin theinsetisfoundinthegraphatangle==4andinthesurroundat=3=4.thetwo theperceivedmotion. motionofthepaperistotheright,thedierenceinerrorvectorsisduetolength,resultingin estimatedmotionvectorsareofthesamelength,eachinthedirectiontowardthegradients ofthelongeredges,andtheprojectionoftheresultingdierencevectoristotheright.ifthe s=t=0:1;in(c),s=t=0:2. thetruemotion.thenoiseisgaussianandthespatialgradientmagnitudeis1.in(a)and(b), betweentheerrorintheinsetandtheerrorinthesurround,andprojecttheresultingvector thetruemotionandthecalculatedmotion.toderivetheslidingmotion,computethedierence onthedominantgradientdirectionintheinset.thelinefromthecenterisinthedirectionof Figure21:Theregionalmotionerrorvectoreld.Thevectorsshownarethedierencesbetween (b) (c) easilyexperienced. oftheinsetforalargerangeofangles(thatis,directionsofeyemovements),theillusionis underfreeviewingconditions,thetriggeringmotionisduetoeyemovements,whichcanbe vectorsoftheinsetandsurroundhasasignicantprojectiononthedominantgradientdirection whichiswhyaclearrelativemotionoftheinsetisseen.whenexperiencingtheouchiillusion approximatedthroughrandom,fronto-paralleltranslations.sincethedierenceinthebias In[34]HelmholtzdescribesanexperimentwiththeZollnerpatternwhichcausesillusionary Weassumethatinadditiontocomputingow,thevisualsystemalsoperformssegmentation, motion.whenthepointofaneedleismadetotraversezollner'spattern(figure13a)hori
27 bandsoccurs.therst,third,fthandseventhblackbandsascend,whilethesecond,fourth, zontallyfromrighttoleft,itsmotionbeingfollowedbytheeye,aperceptionofmotioninthe andsixthdescend;itisjusttheoppositewhenthedirectionofthemotionisreversed. opticalowfromlefttoright.foreachbandtherearetwodierentgradientdirections,i.e., therearetwodierentnormalowcomponentsineachneighborhood.fortheoddbandsthe y-component),andthiscomponentalongthey-axisisperceivedasupwardmotionofthebands twonormalowcomponentsareinthedirectionoftheowanddiagonallytotherightandup. Thustheestimatedowmakesapositiveanglewiththeactualow(thatis,ithasapositive (seefigure22).forevenbandstheestimatedowisbiaseddownward,causingtheperception ofdescentofthebands.similarly,ifthemotionoftheeyeisreversedtheestimatedowhasa Thebiasexplainsthiseectasfollows:Amotionoftheeyesfromrighttoleftgivesriseto negativey-componentintheoddbandsandapositivey-componentintheevenbands,leading toareversalintheperceivedmotionsofthebands. Figure22:Theeyemotiongivesrisetoowu0andnormalowvectorsn1andn2.The theband. 5Theinherentproblem Animportantquestionarises.Istherebiasbecauseofthearchitectureofthevisionsystem? estimatedow,u,hasapositivey-component,whichcausesillusionarymotionupwardin Isthebiasduetothelinearestimationonly?Coulditbecorrectedusingmoresophisticated variables,butanymethodofcompensatingforthebiasrequiresknowledgeofthestatisticsof statisticaltechniques,orcoulditbeavoided? oftenitisnotpossibletoobtainaccurateenoughestimatesofthenoiseparameterstoimprove majorproblem,however,liesintheacquisitionofthestatisticsofthenoise.wearguethat thenoise.inthenoisemodelsconsideredintheprevioussections,thisamountstoknowledge ofthecovariancematrixofthenoise.ifthiswereavailable,inverselterscouldbeappliedto reconstructthegraylevelsignal,andthecorrectedleastsquaresestimatorcouldbeusedto removetheasymptoticbiaswhensolvinglinearsystemsonthebasisofimagederivatives.the Itiswellknownthatlinearestimationisbiasedifthereareerrorsinallthemeasurement thesolution. arefrequencydomainandcorrelationmodels,butcomputationallytheyarenotverydierent. dicultiesoccur,andnoisyestimatesleadtobias.foranextensivediscussionofthestatistics Inallthemodelsthereisastageinwhichsmoothnessassumptionsaremadeandmeasurements withinaregionarecombinedtoobtainmoreexactmeasurements.atthisstagestatistical Thereexistothermodelsforcomputingopticalow.Besidesgradient-basedmodelsthere 26 n 1 n 2 u u
28 ofopticalowestimationsee[8]. A~x=~b,thereiserrorinthevariablesinmatrixAinadditiontotheerrorinthevariablesin hasreceivedalotofattention.theproblematicbiasarisesbecauseinthesystemofequations position,butinformationaboutnoiseratiosisdiculttocompute.itcanbeobtainedonly ables,thatis,theratiosofthetwospatialandtemporalderivativenoisetermsorthenoisein distributed.thismeansthatwehavetoknowtherelativeamountsofnoiseintheerrorvaricallyunbiasedsolutionforsuchsystems,ifthenoisevariablesareindependentandidentically vector~b.thenonlineartotalleastsquaresestimatorhasbeenshowntoprovideanasymptoti- Inrecentyearsthetechniqueoftotalleastsquaresforsolvingsystemsoflinearequations thistechnique:thevarianceislarger.totalleastsquaresisknowntoperformverypoorlyif fromthevariationintheestimatedvariablesovertheimage.thereisanotherproblemwith outliersarepresent,andthesearediculttodetectfromafewmeasurements. acquireenoughdatatocloselyapproximatetheseparameters,butusuallythenoiseparameters longedgesandbarsneedtobedetected,andinthecaseofmotion,discontinuitiesdueto changesindepthanddierentlymovingentitiesneedtobedetectedandthescenesegmented. goodnoisestatisticalotofdataisrequired,sodataneedstobetakenfromlargespatialareas acquiredoveraperiodoftime,butthemodelsusedfortheestimationcanonlybeassumedto holdlocally.thustointegratemoredata,modelsofthesceneneedtobeacquired.specically, Ifthenoiseparametersstayedxedforextendedperiodsoftimeitwouldbepossibleto Whyisitsodiculttoobtainaccurateestimatesofthenoiseparameters?Toacquirea theobjectsbeingviewed,theorientationoftheviewerin3dspace,andthesequenceofeye movementsallhaveinuencesonthenoise.asidefromallthesefactors,inordertoestimate derivatives(ortocomputefouriertransforms)thesystemneedstointerpolate.theaccuracy ofinterpolationcandependincomplexwaysonthepatternofgraylevelsintheimage. othersourcesofnoisebesidessensornoise.thelightingconditions,thephysicalpropertiesof donotstayxedlongenough.sensorcharacteristicsmaystayxed,buttherearemany whichallowittoimprovetheestimates. example,whenviewingastaticpatternwithcontrolledeyemovementsorwhenxatingona thepatternsweakensafterextendedviewing,inparticularwhensubjectsareaskedtoxate. Inthiscasethevisualsystemmayacquirereasonableapproximationsofthenoiseparameters pointintheimageforsometime.thismayexplainwhytheillusionaryperceptioninsomeof Undercontrolledconditionsitmaybepossibletokeepthenoiseparametersxed,for 6DiscussionandSummary Scholarshaveclassiedpatternsasgeometricalopticalillusionsonthebasisofhowtheyare formed(theyarelinedrawings)andtheoreticianshavebeenchallengedtoprovideexplanations forallofthem.thereis,however,noreasontoexpectallillusionaryeectsinvolvingblack largenumberofgeometricalopticalillusions.otherillusionsinthiscategorywhichwehaven't andwhitepatternstobeduetothesamecause. dealtwithare:thecontrasteect(thatis,theoverestimationoflargeextentsnexttosmall guresenclosedbetweenconverginglines(thegurenearertheintersectionofthelinesis intervalseparatedbylinesintosubintervals);theponzoillusions,whichinvolvetwoequal anunlledextent(anexampleistheoppel-kundtillusion anunlledintervalnexttoan unlledextents,thatis,theeectthatalledextentisoverestimatedwhencomparedwith onesandtheunderestimationofsmallextentsnexttolargeones);theillusionsoflledand Inthispaperweadvancedatheoryaboutbiasinthevisualsystemwhichaccountsfora perceptuallyenlarged);thevertical-horizontalillusion,thatis,thephenomenonofjudginga 27
29 areofgreatimportanceinsomeoftheseillusions.thecontrasteectoccursonlyforacertain verticallinetobelongerthanahorizontallineofthesameextent;theframingeect,thatis, (withamaximumofabout2:3),andafterthataninectionoccurs,thatis,forlargerratiosthe oneperceptuallydecreasedandtheinneroneincreased occursuptoarangeofabout1:2 rangeofsizeratios.forexample,thedelboeufillusion twoconcentriccircleswiththeouter atteningofarcs,thatis,theeectthatshortarcsareperceptuallyattened.sizeparameters is,whencertaingureswhichencloseidenticalareasarecompared,andpartoftheboundaryof theoverestimationofframedobjects,andtheeectofoverestimatingunboundedgures,that smallercirclechangesfrombeingoverestimatedtobeingunderestimated.similarly,framing leadstounderestimationofthesizeoftheframedobjectforratiosoftheobjecttotheframe oneismissing,thedimensionwhichistherebyundenedatoneendisoverestimated;andthe closeto1:1. explainedtheillusionnamedafterhimintermsoftheelongationofthearmsofanobtuseangle ascomparedwiththoseofanacuteangle,whilebrentano'sexplanation[2]wasintermsofthe nowadaysasaddinglittletoameredescriptionoftheillusion.thusmuller-lyerhimself theoriesareaimedonlyatonespecicillusionandmostoftheearlytheorieswouldbeseen perceptualenlargementofacuteanglesandthereductionofobtuseangles. Theoriesaboutillusionshavebeenformulatedeversincetheirdiscovery.Manyofthe mathematicalnaturebasedonthegeometryandstatisticsofcapturinglightraysandcomputing ofthebrain(forexample,lateralinhibition).incontrast,thetheoryproposedhereisofa basedonphysicalanalogies,orgeneralobservationsaboutthephysiologyandarchitecture system,articialorbiological.however,onemightndthatourtheoryresemblessomeexisting theirproperties(grayvaluesandspatiotemporalderivatives),andthusitappliestoanyvision toexplaintheworkingsofhumanvisualperception.thesemechanismsareeitherhypothetical, morenoticeinrecentyearsarebasedonaspecicsortofmechanismwhichhasbeensuggested Theorieswhichattempttoexplainabroadspectrumofillusionsandwhichhaveattracted theblurringanddiusingeectofthemediumandtheconstructionoftheeye.theperceived locationsofthelinesareatthepeaksoftheirdistributions;thustwocloselinesinuenceeach fallingontoascreenfromtwoslitsourcesinthestudyofdiractionpatterns.thisisbecauseof theoriesiftheyareputinastatisticalframework. togetheraectoneanother.helikenedthelightcomingintotheeyefromtwolinestothelight other'slocationsandbecomeonewhenthesumoftheirdistributionsformsasinglepeak. Thisleadstoanoverestimationofacuteangles,andprovidesanexplanationoftheZollner, Poggendor,andrelatedillusionsaswellastheMuller-Lyereect.Thediractionamountsto Chiang[6]proposedaretinaltheorythatappliestopatternsinwhichlinesrunningclose data.chiang'smodel,however,ismuchmorelimited.itonlyappliestolines,andconsiders thewholeprocesstotakeplaceintheretina. uncertaintyinthelocationoftheperceivedgraylevelvalues,oritcanbeinterpretedasnoisy inuencethenoisedistributionandthusthebiasperceived,butthereareothernoisesources theyarearelevantsourceofnoise.theparticulareyemovementsmadeinlookingatapattern besideseyemovements,andthisexplainstheexistenceofillusionaryeectsforsomepatterns evenunderxationortachistoscopicviewing. causativefactorinillusions.ourtheoryalsoproposesthateyemovementsplayarolebecause Thereareanumberoftheoriesinwhicheyemovementsareadvancedasanimportant overunlledthanlledelements.inthemuller-lyergurestheeyesmovemorefreelyoverthe healsoexpresseddoubtthattheycouldbethemainsource,asotherillusionsarenotinuenced bythem.carr[5]proposedthattheeyesreacttoaccessorylinesandasaresultpassmoreeasily Helmholtzsuggestedin[34]thatocularmovementsareofimportanceinsomeillusions,but 28
30 gurewithoutgoingnsthanovertheonewithingoingns,andinthepoggendorandzollner guresdeectionsandhesitationsintheeyemovementsareassociatedwiththeintersections distortion. arelinearandrectilinear,horizontalorvertical.whenviewinglineswhichlieothevertical orhorizontal,aneyemovementcorrectionmusttakeplaceandthiscangiverisetoperceptual theeld.virsu[33]suggestedthatatendencytoeyemovements,thatis,instructionsforeye movements,hasaperceptualeect.hesuggeststhattheeyemovementsmostreadilymade Centrationonpartoftheeldcausesanoverestimationofthatpartrelativetotherestof ofthelonglineswiththeobliques.piaget[28]proposeda\lawofrelativecentrations."by \centration"hereferstoakindofcenteringofattentionwhichisverymuchrelatedtoxation. projectionsofthree-dimensionaldisplays[31,32].themostdetailedandmostpopularsuch aftereectsappliedtoillusions[11,18].inthesetheoriesinterferencebetweennearbylines occursbecauseofsatiationinthecortexorlateralinhibitionprocesses.therearealsotheories basedontheassumptionthattheperceptualsysteminterpretsillusionarypatternsasat theoryisduetogregory[12],whoinvokes\sizeconstancyscaling,"whichcanbetriggeredin twodierentways,eitherunconsciouslyorbyhigher-levelawareness,toexplainillusions. Otherfamoustheoriesincludetheorieswhosemainobjectivewastheexplanationofgural aseriousproblemforearlyvisualprocessesandunavoidablyleadstobias.anartifactofthe biasisillusionaryperceptionsinvolvingpatternswherethebiasishighlypronounced.noise intheimagedata thatis,theimagegraylevelanditsspatialandtemporalderivatives causes andtemporalintegrationofdatathatmovingsystemsareconfrontedwith,andduetothe ispresentinanyvisualdata.itisduetothesensingprocess,andinparticularthespatial operationsinvolvedincomputingderivatives,orinestimatingandlocatingcertainfrequency componentsofthesignal.theproblemisthatthenoiseparametersusuallycannotbeestimated Inthispaperwehavediscussedamajorhurdlethatvisionsystemshavetodealwith.Noise anyreferencetotheparticularsettingsthathumanvisionsystemsarefacing,neitherwhat datatobecollected. well,astheychangewiththelightingandviewingconditions,oftentoorapidlytoallowenough smallamountofnoise,andthisallowedustoexplainalargenumberofthegeometricaloptical oftheamountofnoiseandtherelationshipofeyemovementstothenoise.wehaveassumeda thespecicprocessesareandhowtheyareimplementedintheneuralhardware,norinwhich illusions.otherillusionscouldbeexplainediflargeamountsofnoisewerepresent,forexample, usingpsychophysicalandphysiologicalmethodsitshouldbefeasibletoconductinvestigations rangestheuncertaintyinthedatalies.theseareissuesthatneedtobeexamined.inparticular, Thebiashasbeendemonstratedforverygeneralmathematicalmodels.Wehaven'tmade theponzoillusionsorthedelboeufcircles. ofsmoothing,willcausebothcirclesinthemiddleoffigure23atocontract.ifweincrease outercirclewillcontractandtheinnercirclewillexpand,andeventuallythetwocircleswill collapseintoone.thusaratherlargeamountofnoise largeenoughtoaccountformutual theamountofnoisesothatthetwocirclesaecteachotherunderthesmoothing,thenthe inuencesincirclesofsmallersizeratios,butnotlargeenoughtoaccountformutualinuences incirclesoflargersizeratios explainsthendingsinparametricstudiesofthedelboeuf circles.thatis,suchnoisecausesanoverestimationoftheinnercircleforsmallsizeratiosand ConsidertheDelboeufcirclesinFigure23a.Smallamountsofnoise,thatis,smallamounts lineclosertotheintersectionofthelines,causesanexpansionofthisline. seenthatnoiselargeenoughtocausesmoothing,whichaectsoneofthelinesandthevertical anunderestimationforlargerratios.similarly,fortheponzoillusioninfigure23b,itcanbe Theimageformationprocessshouldnotgiverisetolargeamountsofnoise,buttheremay 29
31 informationtoneuronsoflargerreceptiveelds,andviceversa.itisclearthatthehigher-level resolution.thereisevidence[35]thatwithinthehumanvisualsystemtherearehierarchiesof leadsustobelievethattheremustexistsomeprocessesthatcombineinformationfromlocal processesofsegmentationandrecognitionneedinformationfromlargepartsoftheimage.this employingmultiplelevelsofimageresolutioncouldaccountforafairamountofnoise;even neuronsofincreasinglylargereceptiveelds,withneuronsofsmallerreceptiveeldsproviding beothernoiseduetotheprocessinginthevisualsystem.forexample,ascalespaceframework Forexample,tointerpretanyoftheillusionarypatternsinFigure23requiresthatthewhole neighborhoodsintorepresentationsofglobalinformation,andtheseprocessescarryuncertainty. verysmallamountsofnoiseatlowimageresolutiontranslatetolargeamountsofnoiseathigh patternbeunderstoodasjudgmentsofthesizesofdierentpartsofthepatternaremade. Thereisstillmuchresearchtobedonealongtheselines.Visualrecognition,ofcourse,isnot understoodandscalespaceanalysishasbeenlookedatonlywithrespecttothetransformation ofimagefeatures,notwithrespecttoitsutilitarianfunctioninfacilitatingvisualcomputations. Opticalillusionsmighthelpinprovidinginsightsintotheseprocesses. fromphysicsthatthereisuncertaintyinanymeasurement,butvisionresearchershavenot itcausesbias.thusthelocationofalineintheimagemaynotproperlyreectitslocation realizedthattheuncertaintyinthevisualdatahasmajorconsequencesforvisualprocesses; Thendingsofthispaperarealsoofimportanceforpracticalapplications.Itiswellknown Figure23:(a)Delboeufillusion.(b)AversionofthePonzoillusion. (b) inthe3dworldandthelocationoftheintersectionoftwolinesdoesnotcorrespondtoa physicalintersectioninspace.asaconsequence,thereisgreatuncertaintyintheestimation AppendixAExpectedValueoftheLeastSquaresSolution ofcorrespondingfeatures;thisshouldbecleartoallpractitionersofvisionwhoconsiderpoint correspondencesastheirstartingpointforfurthercomputation. Inthissectionasecond-orderTaylorexpansionoftheexpectedvaluesoftheleastsquares solutionforboththeintersectionpoint(section3)andtheow(section4)isgiven. Asthenoiseisassumedtobeindependentandzero-mean,therst-ordertermsaswellasthe IsisthematrixconsistingofthespatialderivativesIxi;Iyi.~CinSection3isthevectorof temporalderivativesitiand~cinsection4isthevectorwhoseelementsareixix0i+iyiy0i. Let~ybethevectortobeestimated,thatis,eithertheintersectionpoint~xortheow~u. second-ordertermsinthenoiseofthetemporalderivatives(orthepositionalparameters)vanish. TheexpectedvalueE(~y)oftheleastsquaressolutionisgivenby E(~y)=E(ItsIs)?1(Its~C) 30
32 Thismeansthatitisonlythenoiseinthespatialderivativeswhichcausesbiasinthemean. TheexpansionatpointsN=0(i.e.,Ixi=Iyi=Iti=0orIxi=Iyi=x0i=y0i=0) canbewrittenase(~y)=~y0+xi Fornotationalsimplicity,wedene 2! Tocomputethepartialderivatives,theexplicittermsofthematrixMare andthetermsof~bare M=264PiI0xi+IxiI0yi+Iyi PiI0xi+Ixi2 ~b=24pii0xi+ixii0ti+iti PiI0yi+IyiI0ti+Iti35 PiI0xi+IxiI0yi+Iyi PiI0yi+Iyi2 375 and~b=264pii0xi+ixi2(x0oi+xoi)+i0xi+ixii0yi+iyi?y0oi+yoi wendtherst-orderandsecond-orderderivativestobe @2~y @xq?1?2m?1"2ixiiyi?m?1"20 Iyi0#M?1"2IxiIyi 00#M?1~b 31
33 SinceweassumeE(I2xi)=E(I2yi),theexpansioncanthusbesimpliedto @I2yiN=02s 2 +Pi(M0?1 "2Ixi0Iyi0 Iyi00#M0?1"2Ixi0Iyi0 Iyi00# +"0Ixi0 Ixi02Iyi0#M0?1"0Ixi0 Ixi02Iyi0#!~y0?M0?1 "2Ixi0Iyi0 +"0Ixi0 wherewehaveunderlinedthetermsthatdonotdependonn(wherenisthenumberof measurementsbeingcombinedinaregion).thesetermswillgiveaconsistent,statistically constantresponse.therestofthetermsdiminishproportionallyto1n.informalexperiments showthatthesetermsbecomenegligibleforn>5,anumberclearlysmallerthanthenumber 0# 2y0i# andthemaintermsintheexpectedvalueoftheintersectionpointe(~x)are E(~x)=~x0?nM0?1~x02s+M0?1"x0i y0i#2s or E(~x)=~x0+nM0?1(~x0?~x0)2s; Iti#; andthesecond-orderderivativesvanish.theexpectedvalueoftheowe(~u)simpliesto E(~u)=~u0?nM0?1~u02s: References [1] [2]F.Brentano.UbereinoptischesParadoxen.Z.Psychologie,3:349{358,
34 [3]J.Canny.Acomputationalapproachtoedgedetection.IEEETransactionsonPattern [4]R.H.S.Carpenter.MovementsoftheEye.Pion,London,1988. [5]H.A.Carr.AnIntroductiontoSpacePerception.Longmans,Green,NewYork,1935. [6]C.Chiang.Anewtheorytoexplaingeometricalillusionsproducedbycrossinglines. AnalysisandMachineIntelligence,8:679{698,1986. [8]C.Fermuller,D.Shulman,andY.Aloimonos.Thestatisticsofopticalow.Computer [7]J.A.Feldman.Fourframessuce:Aprovisionalmodelofvisionandspace.Behavioral andbrainsciences,8:265{313,1985. Percept.Psychophys.,3:174{176,1968. [10]W.Fuller.MeasurementErrorModels.Wiley,NewYork,1987. [12]R.L.Gregory.Distortionofvisualspaceasinappropriateconstancyscaling.Nature, [11]L.Ganz.MechanismoftheF.A.E.'s.Psychol.Rev.,73:128{150,1966. [9]J.Fraser.Anewvisualillusionofdirection.Brit.J.Psychol.,2:307{320,1908. VisionandImageUnderstanding,2001.Toappear. [13]E.Hering.BeitragezurPsychologie,volume1.Engelman,Leipzig,1861. [14]B.K.P.Horn.RobotVision.McGrawHill,NewYork,1986. [15]D.H.HubelandT.N.Wiesel.Integrativeactioninthecat'slateralgeniculatebody.J. 119:678,1963. [18]W.KohlerandH.Wallach.Figuralafter-eects:Aninvestigationofvisualprocesses.Proc. [16] [17] Physiol.(Lond.),155:385{398,1961. [19]A.A.Kundt.UntersuchungenuberAugenmaundoptischeTauschungen.Pogg.Ann., [20]T.Lindeberg.Scale-SpaceTheoryinComputerVision.Kluwer,Boston,1994. Amer.Philos.Soc.,88:269{357,1944. [22]D.MarrandE.C.Hildreth.Atheoryofedgedetection.Proc.RoyalSociety,LondonB, [21]L.MaeiandF.W.Campbell.Neurophysiologicallocalizationofthehorizontaland 207:187{217,1980. verticalcomponentsinman.science,167:386{387, :118{158,1863. [23]F.C.Muller-Lyer.ZurLehrevondenoptischenTauschungen:UberKontrastundKon- [24]H.Munsterberg.DieverschobeneSchachbrettgur.Z.Psychol.,15:184{188,1897. uxion.z.psychologie,9:1{16,
35 [25]H.-H.Nagel.Opticalowestimationandtheinteractionbetweenmeasurementerrorsat [26]J.J.Oppel.Ubergeometrisch-optischeTauschungen.Jahresberichtphys.ver.Frankfurt, adjacentpixelpositions.internationaljournalofcomputervision,15:271{288,1995. [29]J.O.Robinson.ThePsychologyofVisualIllusion.Hutchinson,London,1972. [27]H.Ouchi.JapaneseandGeometricalArt.Dover,NewYork,1977. [28]J.Piaget.LesMechanismesPerceptifs.PressesUniversitairesdeGravee,1961.Translated byg.n.seagrimasthemechanismsofperception,routledge&keganpaul,london, pages37{47,1855. [30]R.W.RodieckandJ.Stone.Analysisofreceptiveeldsofcatretinalganglioncells.J. [31]R.Tausch.OptischeTauschungenalsartizielleEektederGestaltungsprozessevon [32]A.Thiery.UbergeometrischoptischeTauschunge.Phil.Stud.,12:67{126,1896. [33]V.Virsu.Tendenciestoeyemovement,andmisperceptionofcurvature,directionand Neurophysiology,28:833{849, :299{348,1954. GroenundFormkonstanzindernaturlichenRaumwahrnehmung.Psychol.Forsch., [35]S.M.Zeki.AVisionoftheBrain.Blackwell,London,1993. [34]H.L.F.vonHelmholtz.TreatiseonPhysiologicalOptics,volumeIII.Dover,NewYork, 1962.TranslatedfromtheThirdGermanEditionbyJ.P.C.Southall. length.percept.psychophys.,9:339{347,
University of Maryland Fraternity & Sorority Life Spring 2015 Academic Report
University of Maryland Fraternity & Sorority Life Academic Report Academic and Population Statistics Population: # of Students: # of New Members: Avg. Size: Avg. GPA: % of the Undergraduate Population
Logistic Regression. Jia Li. Department of Statistics The Pennsylvania State University. Logistic Regression
Logistic Regression Department of Statistics The Pennsylvania State University Email: [email protected] Logistic Regression Preserve linear classification boundaries. By the Bayes rule: Ĝ(x) = arg max
Normal Distribution as an Approximation to the Binomial Distribution
Chapter 1 Student Lecture Notes 1-1 Normal Distribution as an Approximation to the Binomial Distribution : Goals ONE TWO THREE 2 Review Binomial Probability Distribution applies to a discrete random variable
α α λ α = = λ λ α ψ = = α α α λ λ ψ α = + β = > θ θ β > β β θ θ θ β θ β γ θ β = γ θ > β > γ θ β γ = θ β = θ β = θ β = β θ = β β θ = = = β β θ = + α α α α α = = λ λ λ λ λ λ λ = λ λ α α α α λ ψ + α =
Final Mathematics 5010, Section 1, Fall 2004 Instructor: D.A. Levin
Final Mathematics 51, Section 1, Fall 24 Instructor: D.A. Levin Name YOU MUST SHOW YOUR WORK TO RECEIVE CREDIT. A CORRECT ANSWER WITHOUT SHOWING YOUR REASONING WILL NOT RECEIVE CREDIT. Problem Points Possible
problem arises when only a non-random sample is available differs from censored regression model in that x i is also unobserved
4 Data Issues 4.1 Truncated Regression population model y i = x i β + ε i, ε i N(0, σ 2 ) given a random sample, {y i, x i } N i=1, then OLS is consistent and efficient problem arises when only a non-random
Linear Classification. Volker Tresp Summer 2015
Linear Classification Volker Tresp Summer 2015 1 Classification Classification is the central task of pattern recognition Sensors supply information about an object: to which class do the object belong
STT315 Chapter 4 Random Variables & Probability Distributions KM. Chapter 4.5, 6, 8 Probability Distributions for Continuous Random Variables
Chapter 4.5, 6, 8 Probability Distributions for Continuous Random Variables Discrete vs. continuous random variables Examples of continuous distributions o Uniform o Exponential o Normal Recall: A random
!!! 2014!!2015!NONPROFIT!SALARY!&!STAFFING!REPORT! NEW$YORK$CITY$AREA$ $ $ $ $ $ $
2014 2015NONPROFITSALARY&STAFFINGREPORT NEWYORKCITYAREA 2014NONPROFITSURVEYFINDINGS&2015TRENDS EXECUTIVESUMMARY Thisyear ssalarysurvey201422015confirmsgrowth andchallengesinthesectorduring2014andthestrong
HETEROGENEOUS AGENTS AND AGGREGATE UNCERTAINTY. Daniel Harenberg [email protected]. University of Mannheim. Econ 714, 28.11.
COMPUTING EQUILIBRIUM WITH HETEROGENEOUS AGENTS AND AGGREGATE UNCERTAINTY (BASED ON KRUEGER AND KUBLER, 2004) Daniel Harenberg [email protected] University of Mannheim Econ 714, 28.11.06 Daniel Harenberg
Christfried Webers. Canberra February June 2015
c Statistical Group and College of Engineering and Computer Science Canberra February June (Many figures from C. M. Bishop, "Pattern Recognition and ") 1of 829 c Part VIII Linear Classification 2 Logistic
Triple Integrals in Cylindrical or Spherical Coordinates
Triple Integrals in Clindrical or Spherical Coordinates. Find the volume of the solid ball 2 + 2 + 2. Solution. Let be the ball. We know b #a of the worksheet Triple Integrals that the volume of is given
A General Approach to Variance Estimation under Imputation for Missing Survey Data
A General Approach to Variance Estimation under Imputation for Missing Survey Data J.N.K. Rao Carleton University Ottawa, Canada 1 2 1 Joint work with J.K. Kim at Iowa State University. 2 Workshop on Survey
Linear algebra and the geometry of quadratic equations. Similarity transformations and orthogonal matrices
MATH 30 Differential Equations Spring 006 Linear algebra and the geometry of quadratic equations Similarity transformations and orthogonal matrices First, some things to recall from linear algebra Two
Trigonometry Hard Problems
Solve the problem. This problem is very difficult to understand. Let s see if we can make sense of it. Note that there are multiple interpretations of the problem and that they are all unsatisfactory.
Rotation of Axes 1. Rotation of Axes. At the beginning of Chapter 5 we stated that all equations of the form
Rotation of Axes 1 Rotation of Axes At the beginning of Chapter we stated that all equations of the form Ax + Bx + C + Dx + E + F =0 represented a conic section, which might possibl be degenerate. We saw
Reject Inference in Credit Scoring. Jie-Men Mok
Reject Inference in Credit Scoring Jie-Men Mok BMI paper January 2009 ii Preface In the Master programme of Business Mathematics and Informatics (BMI), it is required to perform research on a business
Root Locus. E(s) K. R(s) C(s) 1 s(s+a) Consider the closed loop transfer function:
Consider the closed loop transfer function: Root Locus R(s) + - E(s) K 1 s(s+a) C(s) How do the poles of the closed-loop system change as a function of the gain K? The closed-loop transfer function is:
Predict Influencers in the Social Network
Predict Influencers in the Social Network Ruishan Liu, Yang Zhao and Liuyu Zhou Email: rliu2, yzhao2, [email protected] Department of Electrical Engineering, Stanford University Abstract Given two persons
Probability for Estimation (review)
Probability for Estimation (review) In general, we want to develop an estimator for systems of the form: x = f x, u + η(t); y = h x + ω(t); ggggg y, ffff x We will primarily focus on discrete time linear
ECON1003: Analysis of Economic Data Fall 2003 Answers to Quiz #2 11:40a.m. 12:25p.m. (45 minutes) Tuesday, October 28, 2003
ECON1003: Analysis of Economic Data Fall 2003 Answers to Quiz #2 11:40a.m. 12:25p.m. (45 minutes) Tuesday, October 28, 2003 1. (4 points) The number of claims for missing baggage for a well-known airline
Computing Euler angles from a rotation matrix
Computing Euler angles from a rotation matrix Gregory G. Slabaugh Abstract This document discusses a simple technique to find all possible Euler angles from a rotation matrix. Determination of Euler angles
CS 431/636 Advanced Rendering Techniques"
CS 431/636 Advanced Rendering Techniques" Dr. David Breen" Korman 105D" Wednesday 6PM 8:50PM" Photon Mapping" 5/2/12" Slide Credits - UC San Diego Goal Efficiently create global illumination images with
Section 5.1 Continuous Random Variables: Introduction
Section 5. Continuous Random Variables: Introduction Not all random variables are discrete. For example:. Waiting times for anything (train, arrival of customer, production of mrna molecule from gene,
Lecture 3: Linear methods for classification
Lecture 3: Linear methods for classification Rafael A. Irizarry and Hector Corrada Bravo February, 2010 Today we describe four specific algorithms useful for classification problems: linear regression,
Phi Theta Kappa Membership. An Honor Earned. Click Here to Learn More. Members & advisor of Omega Nu, Delgado Community College, New Orleans, LA
www.ptk.org Phi Theta Kappa Membership An Honor Earned Click Here to Learn More Members & advisor of Omega Nu, Delgado Community College, New Orleans, LA Phi Theta Kappa Scholar Scholar of largest academic
The Vector or Cross Product
The Vector or ross Product 1 ppendix The Vector or ross Product We saw in ppendix that the dot product of two vectors is a scalar quantity that is a maximum when the two vectors are parallel and is zero
Chapter 10. RC Circuits ISU EE. C.Y. Lee
Chapter 10 RC Circuits Objectives Describe the relationship between current and voltage in an RC circuit Determine impedance and phase angle in a series RC circuit Analyze a series RC circuit Determine
HIPAA Privacy Board Overview
Defense Health Agency Privacy and Civil Liberties Office HIPAA Privacy Board Overview April 30, 2015 1 Objectives The purpose of this presentation is to: Provide an overview of the DHA Privacy and Civil
Pattern Analysis. Logistic Regression. 12. Mai 2009. Joachim Hornegger. Chair of Pattern Recognition Erlangen University
Pattern Analysis Logistic Regression 12. Mai 2009 Joachim Hornegger Chair of Pattern Recognition Erlangen University Pattern Analysis 2 / 43 1 Logistic Regression Posteriors and the Logistic Function Decision
Electromagnetism - Lecture 2. Electric Fields
Electromagnetism - Lecture 2 Electric Fields Review of Vector Calculus Differential form of Gauss s Law Poisson s and Laplace s Equations Solutions of Poisson s Equation Methods of Calculating Electric
Complex Numbers. w = f(z) z. Examples
omple Numbers Geometrical Transformations in the omple Plane For functions of a real variable such as f( sin, g( 2 +2 etc ou are used to illustrating these geometricall, usuall on a cartesian graph. If
THE SINE PRODUCT FORMULA AND THE GAMMA FUNCTION
THE SINE PRODUCT FORMULA AND THE GAMMA FUNCTION ERICA CHAN DECEMBER 2, 2006 Abstract. The function sin is very important in mathematics and has many applications. In addition to its series epansion, it
4.5 Chebyshev Polynomials
230 CHAP. 4 INTERPOLATION AND POLYNOMIAL APPROXIMATION 4.5 Chebyshev Polynomials We now turn our attention to polynomial interpolation for f (x) over [ 1, 1] based on the nodes 1 x 0 < x 1 < < x N 1. Both
Lecture 9. Poles, Zeros & Filters (Lathi 4.10) Effects of Poles & Zeros on Frequency Response (1) Effects of Poles & Zeros on Frequency Response (3)
Effects of Poles & Zeros on Frequency Response (1) Consider a general system transfer function: zeros at z1, z2,..., zn Lecture 9 Poles, Zeros & Filters (Lathi 4.10) The value of the transfer function
Section 5.4 More Trigonometric Graphs. Graphs of the Tangent, Cotangent, Secant, and Cosecant Function
Section 5. More Trigonometric Graphs Graphs of the Tangent, Cotangent, Secant, and Cosecant Function 1 REMARK: Many curves have a U shape near zero. For example, notice that the functions secx and x +
Forces & Magnetic Dipoles. r r τ = μ B r
Foces & Magnetic Dipoles x θ F θ F. = AI τ = U = Fist electic moto invented by Faaday, 1821 Wie with cuent flow (in cup of Hg) otates aound a a magnet Faaday s moto Wie with cuent otates aound a Pemanent
How to Obtain Financial Success in 2009?
SMERescueLoansProgram APublic PrivateProgramtoSustainandCreateEmploymentthroughIncentivesforPrivateRescueLendingto SmallandMiddle marketenterprises www.smerescueloans.com Summary Asfinanciallyimpairedbankshaveretrenchedfromtraditionalsecuredlendingtosmalland
Probabilistic Models for Big Data. Alex Davies and Roger Frigola University of Cambridge 13th February 2014
Probabilistic Models for Big Data Alex Davies and Roger Frigola University of Cambridge 13th February 2014 The State of Big Data Why probabilistic models for Big Data? 1. If you don t have to worry about
FLUID MECHANICS. TUTORIAL No.7 FLUID FORCES. When you have completed this tutorial you should be able to. Solve forces due to pressure difference.
FLUID MECHANICS TUTORIAL No.7 FLUID FORCES When you have completed this tutorial you should be able to Solve forces due to pressure difference. Solve problems due to momentum changes. Solve problems involving
Lecture 4: BK inequality 27th August and 6th September, 2007
CSL866: Percolation and Random Graphs IIT Delhi Amitabha Bagchi Scribe: Arindam Pal Lecture 4: BK inequality 27th August and 6th September, 2007 4. Preliminaries The FKG inequality allows us to lower bound
Classification Problems
Classification Read Chapter 4 in the text by Bishop, except omit Sections 4.1.6, 4.1.7, 4.2.4, 4.3.3, 4.3.5, 4.3.6, 4.4, and 4.5. Also, review sections 1.5.1, 1.5.2, 1.5.3, and 1.5.4. Classification Problems
A hidden Markov model for criminal behaviour classification
RSS2004 p.1/19 A hidden Markov model for criminal behaviour classification Francesco Bartolucci, Institute of economic sciences, Urbino University, Italy. Fulvia Pennoni, Department of Statistics, University
Online Supplementary Material
Online Supplementary Material The Supplementary Material includes 1. An alternative investment goal for fiscal capacity. 2. The relaxation of the monopoly assumption in favor of an oligopoly market. 3.
www.sakshieducation.com
LENGTH OF THE PERPENDICULAR FROM A POINT TO A STRAIGHT LINE AND DISTANCE BETWEEN TWO PAPALLEL LINES THEOREM The perpendicular distance from a point P(x 1, y 1 ) to the line ax + by + c 0 is ax1+ by1+ c
Capital Budgeting: Decision. Example. Net Present Value (NPV) FINC 3630 Yost
Capital Budgeting: Decision Criteria Example Consider a firm with two projects, A and B, each with the following cash flows and a 10 percent cost of capital: Project A Project B Year Cash Flows Cash Flows
EDUCATION: Ph.D. May 2017 (expected) West Virginia University, Morgantown, WV Psychology (Behavioral Neuroscience)
Sara R. Nass Graduate Research Associate West Virginia University Department of Psychology PO Box 6040 Morgantown, WV 26506-6040 http://community.wvu.edu/~sgk008/nass.html e-mail: [email protected] EDUCATION:
Introduction to the Monte Carlo method
Some history Simple applications Radiation transport modelling Flux and Dose calculations Variance reduction Easy Monte Carlo Pioneers of the Monte Carlo Simulation Method: Stanisław Ulam (1909 1984) Stanislaw
Solving Cubic Polynomials
Solving Cubic Polynomials 1.1 The general solution to the quadratic equation There are four steps to finding the zeroes of a quadratic polynomial. 1. First divide by the leading term, making the polynomial
Eligibility to Serve as a Principal Investigator for Research Involving Human Subjects
Eligibility to Serve as a Principal Investigator for Research Involving Human Subjects Statement of Policy Tenured or tenure track Purdue faculty, including emeriti, who are certified by Purdue University
Why Use Net Present Value? The Payback Period Method The Discounted Payback Period Method The Average Accounting Return Method The Internal Rate of
1 Why Use Net Present Value? The Payback Period Method The Discounted Payback Period Method The Average Accounting Return Method The Internal Rate of Return Problems with the IRR Approach The Profitability
Today in Physics 217: the method of images
Today in Physics 17: the method of images Solving the Laplace and Poisson euations by sleight of hand Introduction to the method of images Caveats Example: a point charge and a grounded conducting sphere
Dongfeng Li. Autumn 2010
Autumn 2010 Chapter Contents Some statistics background; ; Comparing means and proportions; variance. Students should master the basic concepts, descriptive statistics measures and graphs, basic hypothesis
9231 FURTHER MATHEMATICS
CAMBRIDGE INTERNATIONAL EXAMINATIONS GCE Advanced Subsidiary Level and GCE Advanced Level MARK SCHEME for the October/November 2012 series 9231 FURTHER MATHEMATICS 9231/21 Paper 2, maximum raw mark 100
Basics of Statistical Machine Learning
CS761 Spring 2013 Advanced Machine Learning Basics of Statistical Machine Learning Lecturer: Xiaojin Zhu [email protected] Modern machine learning is rooted in statistics. You will find many familiar
How to Graph Trigonometric Functions
How to Graph Trigonometric Functions This handout includes instructions for graphing processes of basic, amplitude shifts, horizontal shifts, and vertical shifts of trigonometric functions. The Unit Circle
5 Directed acyclic graphs
5 Directed acyclic graphs (5.1) Introduction In many statistical studies we have prior knowledge about a temporal or causal ordering of the variables. In this chapter we will use directed graphs to incorporate
Section 9.1 Vectors in Two Dimensions
Section 9.1 Vectors in Two Dimensions Geometric Description of Vectors A vector in the plane is a line segment with an assigned direction. We sketch a vector as shown in the first Figure below with an
THE COMPLEX EXPONENTIAL FUNCTION
Math 307 THE COMPLEX EXPONENTIAL FUNCTION (These notes assume you are already familiar with the basic properties of complex numbers.) We make the following definition e iθ = cos θ + i sin θ. (1) This formula
Ethics, Privilege, and Practical Issues in Cloud Computing, Privacy, and Data Protection: HIPAA February 13, 2015
Ethics, Privilege, and Practical Issues in Cloud Computing, Privacy, and Data Protection: HIPAA February 13, 2015 Katherine M. Layman Cozen O Connor 1900 Market Street Philadelphia, PA 19103 (215) 665-2746
Low Default Portfolio (LDP) modelling
Low Default Portfolio (LDP) modelling Probability of Default (PD) Calibration Conundrum 3 th August 213 Introductions Thomas Clifford Alexander Marianski Krisztian Sebestyen Tom is a Senior Manager in
Path tracing everything. D.A. Forsyth
Path tracing everything D.A. Forsyth The Rendering Equation- 1 We can now write L o (x, ω o )=L e (x, ω o )+ Ω Angle between normal and incoming direction ρ bd (x, ω o, ω i )L i (x, ω i ) cos θ i dω i
MAT 1341: REVIEW II SANGHOON BAEK
MAT 1341: REVIEW II SANGHOON BAEK 1. Projections and Cross Product 1.1. Projections. Definition 1.1. Given a vector u, the rectangular (or perpendicular or orthogonal) components are two vectors u 1 and
Dot product and vector projections (Sect. 12.3) There are two main ways to introduce the dot product
Dot product and vector projections (Sect. 12.3) Two definitions for the dot product. Geometric definition of dot product. Orthogonal vectors. Dot product and orthogonal projections. Properties of the dot
Polarized 3D. Achuta Kadambi, MIT Media Lab. Joint work with Vahe Taamazyan, Boxin Shi, and Ramesh Raskar. MIT Media Lab 2
Polarized 3D Achuta Kadambi, MIT Media Lab Joint work with Vahe Taamazyan, Boxin Shi, and Ramesh Raskar MIT Media Lab 2 MIT Media Lab 3 Microsoft Kinect v2 MIT Media Lab 4 Microsoft Kinect v2 MIT Media
Solution: F = kx is Hooke s law for a mass and spring system. Angular frequency of this system is: k m therefore, k
Physics 1C Midterm 1 Summer Session II, 2011 Solutions 1. If F = kx, then k m is (a) A (b) ω (c) ω 2 (d) Aω (e) A 2 ω Solution: F = kx is Hooke s law for a mass and spring system. Angular frequency of
Cross product and determinants (Sect. 12.4) Two main ways to introduce the cross product
Cross product and determinants (Sect. 12.4) Two main ways to introduce the cross product Geometrical definition Properties Expression in components. Definition in components Properties Geometrical expression.
Derive 5: The Easiest... Just Got Better!
Liverpool John Moores University, 1-15 July 000 Derive 5: The Easiest... Just Got Better! Michel Beaudin École de Technologie Supérieure, Canada Email; [email protected] 1. Introduction Engineering
Recursive Estimation
Recursive Estimation Raffaello D Andrea Spring 04 Problem Set : Bayes Theorem and Bayesian Tracking Last updated: March 8, 05 Notes: Notation: Unlessotherwisenoted,x, y,andz denoterandomvariables, f x
2.1 Three Dimensional Curves and Surfaces
. Three Dimensional Curves and Surfaces.. Parametric Equation of a Line An line in two- or three-dimensional space can be uniquel specified b a point on the line and a vector parallel to the line. The
Section 9.5: Equations of Lines and Planes
Lines in 3D Space Section 9.5: Equations of Lines and Planes Practice HW from Stewart Textbook (not to hand in) p. 673 # 3-5 odd, 2-37 odd, 4, 47 Consider the line L through the point P = ( x, y, ) that
OHIO REGION PHI THETA KAPPA 2012-13
OHIO REGION PHI THETA KAPPA REGION HALLMARK AWARDS HONORS IN ACTION HALLMARK WINNER Alpha Rho Epsilon Columbus State Community College HONORS IN ACTION HALLMARK FIRST RUNNER-UP Washington State Community
1 Inner Products and Norms on Real Vector Spaces
Math 373: Principles Techniques of Applied Mathematics Spring 29 The 2 Inner Product 1 Inner Products Norms on Real Vector Spaces Recall that an inner product on a real vector space V is a function from
An Introduction to Basic Statistics and Probability
An Introduction to Basic Statistics and Probability Shenek Heyward NCSU An Introduction to Basic Statistics and Probability p. 1/4 Outline Basic probability concepts Conditional probability Discrete Random
Lecture 8: More Continuous Random Variables
Lecture 8: More Continuous Random Variables 26 September 2005 Last time: the eponential. Going from saying the density e λ, to f() λe λ, to the CDF F () e λ. Pictures of the pdf and CDF. Today: the Gaussian
V Estrella. Victoria Estrella. Dr. Gail F. Buck. CSA 562: Today s College Students. 7, April, 2007. Assignment: Final Report
1 Napa Valley College Transfer Students are Missing Out on Greek Life. Victoria Estrella Dr. Gail F. Buck CSA 562: Today s College Students 7, April, 2007 Assignment: Final Report 2 Abstract This study
Given three vectors A, B, andc. We list three products with formula (A B) C = B(A C) A(B C); A (B C) =B(A C) C(A B);
1.1.4. Prouct of three vectors. Given three vectors A, B, anc. We list three proucts with formula (A B) C = B(A C) A(B C); A (B C) =B(A C) C(A B); a 1 a 2 a 3 (A B) C = b 1 b 2 b 3 c 1 c 2 c 3 where the
Quantum Algorithms. Peter Høyer
Quantum Algorithms Peter Høyer University of Calgary CSSQI 2015 Canadian Summer School on Quantum Information Toronto, August 11, 2015 QUERY MODEL x1 x2 xn x1 0 1 0 0 0 1 0 1 0 0 0 0 xn You can ask questions
