Video-BasedOnline FaceRecognitionUsing Identity Surfaces

YongminLi, ShaogangGongandHeatherLiddellDepartmentof ComputerScience,

QueenMary, Universityof London,LondonE1 4NS,UK

Email:�yongmin,sgg,heather� @dcs.qmw.ac.uk

Abstract

Recognising facesacrossmultiple views is more chal-lenging than that from a fixedview becauseof the severenon-linearity causedby rotation in depth, self-occlusion,self-shading, and change of illumination. The problemcan be related to the problem of modelling the spatio-temporal dynamicsof moving facesfrom video input forunconstrainedlive facerecognition. Bothproblemsremainlargely under-developed.To addresstheproblems,a novelapproach is presentedin this paper. A multi-view dynamicfacemodelis designedto extract the shape-and-pose-freetexturepatternsof faces.Themodelprovidesa precisecor-respondenceto the taskof recognition sincethe 3D shapeinformation is usedto warp the multi-view facesonto themodelmeanshapein frontal-view. The identity surfaceofeach subjectis constructedin a discriminantfeature spacefrom a sparse set of face texture patterns,or more prac-tically, from one or more learning sequencescontainingthe face of the subject. Insteadof matching templatesorestimatingmulti-modaldensityfunctions,facerecognitioncanbeperformedbycomputingthepatterndistancesto theidentity surfacesor trajectorydistancesbetweentheobjectandmodeltrajectories.Experimentalresultsdepictthatthisapproach providesanaccuraterecognitionratewhileusingtrajectory distancesachievesa more robust performancesincethe trajectoriesencodethe spatio-temporal informa-tion and contain accumulatedevidenceabout the movingfacesin a videoinput.

1 Intr oduction

The issueof facerecognitionhasbeenextensively ad-dressedover the past decade. Various models and ap-proacheshave beenproposedaiming to solve the problemunderdifferentassumptionsandconstraints.Amongthem,theeigenfaceapproachproposedin [16, 17] usesPrincipalComponentAnalysis(PCA) to codefaceimagesandcap-

ture facefeatures. In [19], facerecognitionis performedby ElasticGraphmatchingbasedon aGaborwavelettrans-form. TheActiveShapeModel (ASM) andActiveAppear-anceModel (AAM) capturingboth shapeand shape-freegrey-level appearanceof faceimageshave beenappliedtofacemodellingandrecognition[4, 3]. Thesemethodshavebeenmostly applied in frontal-view or near frontal-viewfacerecognition.

However, recognisingfaceswith largeposevariationismorechallengingbecauseof theseverenon-linearitycausedby rotationin depth,self-occlusion,self-shadingandillumi-nationchange.Theeigenfacemethodhasbeenextendedtoview-basedandmodulareigenspacesintendedfor recognis-ing facesundervaryingviews[13]. Li et al. [10] presenteda view-basedpiece-wiseSVM (SupportVector Machine)model of the facespace. Cooteset al. [5] proposedtheview-basedActiveAppearanceModelswhichemploy threemodelsfor profile,half-profileandfrontalviews. But thedi-visionof thefacespacein thesemethodsis ratherarbitrary,ad hoc andoften coarse.Both ASM andAAM have beenextendedto nonlinearcasesacrossviews basedon KernelPrincipalComponentAnalysis(KPCA)[14]. Thesenonlin-earmodelsaimedatcorrespondingdynamicappearancesofboth shapeand texture acrossviews, but the computationinvolvedis ratherintensive.

Also, in mostof the previous work, the basicmethod-ology adoptedfor recognitionis largely basedon match-ing staticfaceimagepatternsin a featurespace.Psychol-ogy andphysiologyresearchdepictsthat thehumanvisionsystem’s ability to recogniseanimatedfacesis betterthanthat on randomlyorderedstill faceimages(i.e. the sameset of images,but displayedin randomorder without thetemporalcontext of moving faces). Knight and Johnston[9] arguedthat recognitionof famousfacesshown in pho-tographicnegativescould be significantlyenhancedwhenthe faceswereshown moving ratherthanstatic. Bruceetal. [1, 2] extendedthis result to other conditionswhererecognitionis madedifficult, e.g. by thresholdingthe im-agesor showing themin blurredor pixellatedformats. In

a computervision system,whenfacesaretrackedconsecu-tively, notonly moreinformationof thosefacesfrom differ-entviews but alsothe spatio-temporalconnectionbetweenthosefacescan be obtained[7]. Yamaguchiet al. [20]presenteda methodfor facerecognitionfrom sequencesbybuilding a subspacefor the detectedfaceson thegivense-quenceandthenmatchingthesubspacewith prototypesub-spaces.Gonget al. [8] introducedan approachthat usesPartially RecurrentNeuralNetworks(PRNNs)to recognisetemporalsignaturesof faces. Other work of recognisingfacesfrom video sequencesinclude[6, 15]. Nevertheless,theissueof recognisingthedynamicsof humanfacesunderaspatio-temporalcontext remainslargelyunresolved.

In thispaper, wepresentanovelapproachto video-basedfacerecognition. In the registrationstage,an identity sur-face for eachsubjectto be recognisedis constructedin adiscriminantfeaturespacefrom one or more learningse-quences, while in run-time, recognitionis performedbycomputingthe patterndistancesto the identity surfacesorthetrajectorydistancesbetweentheobjecttrajectoryandasetof modeltrajectorieswhich encodethespatio-temporalinformationof a moving faceover time. A multi-view dy-namicfacemodelisdesignedtoextracttheshape-and-pose-free texture patternsfrom sequencesfor accurateacross-view registrationin bothtrainingandrun-time.

Theremainingpartof this paperis arrangedasfollows:Themulti-view dynamicfacemodelis describedin Section2. Identity surfacesynthesis,objectandmodel trajectoryconstruction,anddynamicfacerecognitionarepresentedinSection3. Conclusionsaredrawn in Section4.

2 Multi-V iew Dynamic FaceModel

Our multi-view dynamicfacemodel [12] consistsof asparse3D PointDistribution Model (PDM) [4] learntfrom2D imagesin differentviews,ashape-and-pose-freetexturemodel,andanaffine geometricalmodelwhich controlstherotation,scaleandtranslationof faces.

The3D shapevectorof a faceis estimatedfrom a setof2D faceimagesin differentviews, i.e. given a setof 2Dfaceimageswith known poseand2D positionsof theland-marks,the 3D shapevectorcanbe estimatedusing linearregression.To decouplethecovariancebetweenshapeandtexture, a faceimagefitted by the shapemodel is warpedto themeanshapeat frontal view (with ��� in both tilt andyaw), obtainingashape-and-pose-freetexturepattern.Thisis implementedby forming a triangulationfrom the land-marksandemployingapiece-wiseaffinetransformationbe-tweeneachtriangle pair. By warping to the meanshape,oneobtainstheshape-freetextureof thegivenfaceimage.Furthermore,by warping to the frontal view, a pose-freetexturerepresentationis achieved. We appliedPCA on the3D shapepatternsandshape-and-pose-freetexturepatterns

respectively to obtaina low dimensionalstatisticalmodel.

Figure1 showsthesamplefaceimagesusedto constructthe model, the landmarkslabelledon eachimage,the 3Dshapeestimatedfrom theselabelledfaceimages,and theextractedshape-and-pose-freetexturepatterns.

Figure 1. Multi-vie w dynamic face model.From top to bottom are sample training faceimages (fir st row), the landmarks labelled onthe images (second row), the estimated 3Dshape rotating from ������� to ����� in yaw andwith tilt fix ed on ��� , and the extracted shape-and-pose-free texture patterns.

Basedon the analysisabove, a facepatterncanbe rep-resentedin thefollowing way. First, the3D shapemodelisfitted to a givenimageor videosequencecontainingfaces.A SupportVectorMachinebasedposeestimationmethod[10] is employed to obtain the tilt and yaw anglesof thefacein theimage.Thenthefacetextureis warpedontothemeanshapeof the3D PDM modelin frontal view. Finally,by addingparameterscontrolling pose,shift andscale,thecompleteparametersetof the dynamicmodel for a givenfacepatternis � ����������������������������� ��!#"%$ where � is theshapeparameter, � is the texture parameter, �������&" is posein tilt andyaw, �'���������(" is thetranslationof thecentroidoftheface,and ! is its scale.More detailsof modelconstruc-tion andfitting aredescribedin [12].

Theshape-and-pose-freetexturepatternsobtainedfrommodelfitting areadoptedfor facerecognition.In ourexper-iments,we alsotried to usetheshapepatternsfor recogni-tion, however, the performancewasnot asgoodasthat ofusingtextures.

3 RecognisingFacesUsing Identity Surfaces

For appearancebasedmodels,it is morechallengingtorecognisefacesacrossmultiple views since(1) theappear-anceof differentpeoplefrom a sameview is moresimilarthanthatof onepersonfrom differentviews,and(2) thero-tationin depth,self-occlusionandself-shadingbringseverenon-linearityto the task. Thereforeemphasisingthe vari-ation betweendifferentsubjectsandsuppressingthe vari-ation within eachsubjectat the sametime, arethe key is-suesof multi-view facerecognition.Themostwidely usedtechniquesfor facerecognitionincludecomputingthe Eu-clideanor Mahalanobisdistanceto a meanvectoror a tem-plate of a face class, and estimatingthe density of pat-ternsusingmulti-modalmodels.However, thesetechniquescannotprovide straightforwardsolutionsto theproblemofmulti-view facerecognition.This situationis illustratedinFigure2 wherethePCApatternsof multi-view faceimagesfrom four subjectsaremingledtogether. Moreprecisely, thevariationbetweendifferentsubjectsis overwhelmedby thatfrom posechange.

−4000 −2000 0 2000 4000 6000 8000−2500

−2000

−1500

−1000

−500

0

500

1000

1500

2000

2500

PCA1

PC

A2

Figure 2. Distrib ution of multi-vie w face pat-terns from four subjects represented in twoPCA dimensions. There are mainl y two kindsof variation among these patterns: variationfrom diff erent subjects and variation frompose chang e. Unfor tunatel y the former isoverwhelmed by the latter .

To addressthisproblem,weproposeanapproachto con-struct identity surfacesin a discriminantfeaturespaceformulti-view facerecognitionwheretheposeinformationoffacesis explicitly used. Eachsubjectto be recognisedisrepresentedby a uniquehypersurfacebasedon poseinfor-mation. In otherwords,thetwo basiscoordinatesstandforthe headpose: tilt andyaw, andthe othercoordinatesare

usedto representthediscriminantfeaturepatternsof faces.For eachpairof tilt andyaw, thereis oneunique“point” fora faceclass.Thedistributionof all these“points” of asamefaceforms a hypersurfacein this featurespace. We callthis surfacean identity surface. Facerecognitioncanthenbe performedby computingand comparingthe distancesbetweena givenpatternandasetof identitysurfaces.

3.1 SynthesisingIdentity Surfacesof Faces

If sufficientpatternsof afacein differentviewsareavail-able,theidentitysurfaceof this facecanbeconstructedpre-cisely. However, wedonotpresumesuchastrict condition.In thiswork, wedevelopamethodto synthesisethe identitysurfaceof a subjectfrom a small sampleof facepatternswhich sparselycover theview sphere.Thebasicideais toapproximatethe identity surfaceusinga setof )+* planesseparatedby a numberof )-, predefinedviews. Theprob-lemcanbeformally definedasfollows:

Suppose����� aretilt andyaw respectively, . is the dis-criminant featurevector of a facepattern,e.g. the KDA1 vector. ���0/21#���#/213"3�4�5�0/76����#/768"2�:9;9<9;�8���0/7=�>?���#/7=�>@" areprede-fined views which separatethe view planeinto ) * pieces.Oneachof these) * pieces,the identitysurfaceis approxi-matedby a plane

.-�BAC�DFE&�GH� (1)

Supposethe IFJ samplepatternscoveredby the K th planeare �5�LJM1#����JN1#��.�JN18"3�8���OJ;6�����J;6���.�J;64"2�:9;9<9;�8���OJMPRQ:����JMPRQ2��.�JMPQ�" ,thenoneminimises

S �=UTVJ

P QVWYX ACJ��OJ W ZE&J���J W H��J[�F.�J W X

6 (2)

subjectto \]A J � /3^ FE J � /7^ H� J �BA�_8� /7^ ZE`_:� /7^ H�a_b �B�0�:c��:9;9<9;��)d,��planesK��%e intersectat ���0/3^?���#/3^#"29 (3)

This is a quadraticoptimisation problem which can besolvedusingtheinteriorpoint method[18].

For an unknown facepattern ���U��� ��.�f?" where .�f is theKDA vectorand ����� arethe posein tilt andyaw, onecancomputethe distanceto oneof the identity surfacesastheEuclideandistancebetween.�f andthecorrespondingpointon the identitysurface.

�d� X .�fG�g. X (4)

where. is givenby (1).

1KernelDiscriminantAnalysis,a nonlinearapproachto maximisetheinter-classvariationandminimisethewithin-classvariation[11].

3.2 Video-BasedOnline FaceRecognition

Whenafaceis trackedcontinuouslyin avideosequence,not only additionalinformationin differentviewsandfrommultiple framesis obtained,but also the underlying dy-namic characteristicsof the face can be capturedby thespatio-temporalcontinuity, variation, and distribution offacepatterns.

For computerbasedvision systems,formulating andmodelling the psychologicaland physiologicaldynamicsdescribedin Section1 is still an unsolvedproblem. How-ever, significantimprovementin termsof recognitionaccu-racy androbustnessmaystill beachievedwhenthespatio-temporalinformation is modelledin a rather straightfor-ward way, e.g. simply accumulatingthe discriminantevi-dencewith the spatio-temporalorderencodedin an inputsequence.Basedon this idea,we formulatethe followingapproachto video-basedfacerecognition.

Figure 3. Identity surfaces for face recogni-tion. When a face is trac ked contin uousl yfrom a video input, a robust recognition canbe perf ormed by matc hing the object trajec-tor y to a set of model trajectories.

As shown in Figure 3, when a face is detectedandtracked in an input video sequence,oneobtainsthe objecttrajectoryof thefacein thefeaturespace.Also, its projec-tion oneachof the identitysurfacewith thesameposesandtemporalorderformsamodeltrajectoryof thespecificsub-ject. It canberegardedastheidealtrajectoryof thissubjectencodedby thesamespatio-temporalinformation(posein-formationandtemporalorderfrom the videosequence)asthe tracked face. Thenfacerecognitioncanbe carriedoutby matchingtheobjecttrajectorywith asetof modeltrajec-tories. Comparedto facerecognitionon staticimages,thisapproachcanbe more robust andaccurate.For example,

it is difficult to decidewhetherthe pattern h in Figure3belongsto subjectA or B for a singlepattern,however, ifwe know that h is trackedalongtheobjecttrajectory, it ismuchclearthatit is morelikely to besubjectA thanB.

3.3 Constructing Identity Surfacesfr omLearningSequences

Beforerecognitionis carriedout, a faceclassshouldberegisteredto thesystemby oneor morelearningsequencescontainingthe faceof the subject. For example,we canrecordasmallvideoclip of thissubjectwhile he/sherotatestheheadin front of acamera.After applyingthemulti-viewdynamicfacemodeldescribedin Section2 on thevideose-quence,weobtainasetof facepatternsof thissubject.Thenthesepatternsarestoredto constructthe identitysurfaceofthissubjectand,if necessary, to train (or re-train)theKDA.

To simplify computation,normallywe do not useall thepatternsof eachsubjectto train the KDA sincethe sizeofthekernelmatrix is directly relatedto thenumberof train-ing examples.A pragmaticway to selecttheKDA trainingpatternsis to factor-samplethepatternsfrom thetrainingse-quencessothattheresultpatternsuniformly cover theviewsphere.

After theKDA training,all facepatternscanbeprojectedontothefeaturespacespannedby thesignificantKDA basevectors. Thenthe methoddescribedin Section3.1 canbeemployedto constructthe identitysurfaces.

3.4 Object Trajectory and Model Trajectories

In the recognitionstage,we apply the samemulti-viewdynamicfacemodelon a novel sequencecontainingfacesto berecognised,thenanobjecttrajectorycanbeobtainedby projectingthefacepatternsinto theKDA featurespace.On theotherhand,accordingto theposeinformationof thefacepatterns,it is easyto build themodeltrajectoryon theidentitysurfaceof eachsubjectusingthe sameposeinfor-mationandtemporalorderof the object trajectory. Thosetwo kinds of trajectories,i.e. object and model trajecto-ries,encodethespatio-temporalinformationof thetrackedface. And finally, the recognitionproblemcan be solvedby matchingtheobjecttrajectoryto a setof identity modeltrajectories.

A preliminary realisation of this approachis imple-mentedby computing the trajectory distances,i.e. theweightedsumof thedistancesbetweenpatternson theob-jecttrajectoryandtheircorrespondingpatternsonthemodeltrajectories.At frame i of a sequence,we definethe dis-tancebetweenthe object trajectoryandan identity modeltrajectoryj as:

� W � kV J<l�1`m J'� W J (5)

Figure 4. Video-base multi-vie w face recognition. From top to bottom, sample images from a testsequence with an inter val of 10 frames, images fitted by the multi-vie w dynamic face model, and theshape-and-pose-free texture patterns.

where � W J , the patterndistancebetweenthe facepatterncapturedin the K th frameandtheidentitysurfaceof the j thsubject,is computedfrom (4), and m J is theweighton thisdistance.Considerationson determiningm J include: con-fidenceof modelfitting, variationfrom thepreviousframe,andview of the facepattern(e.g. profile facepatternsareweightedlower sincethey carry lessdiscriminantinforma-tion). Finally, recognitioncanbeperformedby:

Kn�d�po�!4q?jrK%s PW l�1 � W (6)

3.5 Experiments

We demonstratethe performanceof this approachon asmall scalemulti-view facerecognitionproblem. Twelvesequences,eachfrom a set of 12 subjects,were usedastraining sequencesto constructthe identity surfaces. Thenumberof framescontainedin eachsequencevariesfrom40to 140.Only 10KDA dimensionswereusedto constructthe identity surfaces. Thenrecognitionwasperformedonnew test sequencesof thesesubjects. Figure4 shows thesampleimagesfittedby ourmulti-view dynamicmodelandthewarpedshape-and-pose-freetexturepatternsfrom atestsequence.Theobjectandmodeltrajectories(in thefirst twoKDA dimensions)areshown in Figure5. The patterndis-tancesfrom the identity surfacesin eachindividual frameareshown in the left sideof Figure6, while the trajectorydistancesshown in the right side. Theseresultsdepictthata more robust performanceis achieved when recognitionis carriedout usingthe trajectorydistanceswhich includetheaccumulatedevidenceover time thoughthepatterndis-tancesto the identity surfacesin eachindividual frameal-readyprovidesa sufficient recognitionaccuracy.

4 Conclusions

Most of the previous researchin face recognition ismainly on frontal-view (or near frontal-view) and fromstaticimages.In this work, anapproachto video-basedon-line facerecognitionis presented.Thecontributionsof thiswork include:

1. A multi-view dynamicalfacemodelis adoptedto ex-tract theshape-and-pose-freetexturepatternsof facesfrom a videoinput. Themodelprovidesa precisecor-respondencefor recognitionsince3D shapeinforma-tion is usedto warp the texture patternsto the modelmeanshapein frontal view.

2. Insteadof matching templatesor estimatingmulti-modal density, the identity surfacesof face classesareconstructedin a discriminantfeaturespace.Thenrecognitionis performedby computingthepatterndis-tancesto the identitysurfaces.

3. A more robust performancecan be achieved by per-forming recognitiondynamicallyon videosequences.The identity surfacescan be constructedfrom learn-ing sequences.Trajectorydistancesbetweenthe ob-ject andmodel trajectories,which encodethe spatio-temporalinformationof a moving face,is computedfor recognition.

For visual interactionandhuman-computerinteraction,theproblemof facerecognitioninvolvesmorethanmatch-ing staticimages.At a low-level, thefacedynamicsshouldbe accommodatedin a consistentspatio-temporalcontextwheretheunderlyingvariationswith respectto changesinidentity, view, scale,position, illumination, andocclusionare integratedtogether. At a higher level, more sophis-ticatedbehaviour models,including individual-dependent

0 10 20 30 40 50 60 70 80 90 100−0.25

−0.2

−0.15

−0.1

−0.05

0

0.05

0.1

0.15

0.2

0.25

Frame

KD

A1

0 10 20 30 40 50 60 70 80 90 100−0.25

−0.2

−0.15

−0.1

−0.05

0

0.05

0.1

0.15

0.2

0.25

Frame

KD

A2

Figure 5. The object and model trajectories in the fir st two KDA dimensions. The object trajecto-ries are the solid lines with dots denoting the face patterns in each frame . The other s are modeltrajectories where the ones from the ground-truth subject highlighted with solid lines.

0 10 20 30 40 50 60 70 80 90 1000

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

Frame

Pat

tern

Dis

tanc

e

0 10 20 30 40 50 60 70 80 90 1000

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

Frame

Tra

ject

ory

Dis

tanc

e

Figure 6. Pattern distances and trajector y distances. The ground-truth subject is highlighted withsolid lines. By using KDA and identity surfaces, the pattern distances can alread y give an accu-rate result. However, the trajector y distances provide a more robust perf ormance , especiall y itsaccum ulated effects (i.e. discriminative ability) over time .

andindividual-independentmodels,maysuperviseandco-operatewith all thelow level modules.

In this paper, we highlightedthe natureof the problemandshowedthepotentialof modellingfacedynamicsasaneffective meansto facerecognition.Nevertheless,someofthe implementationsuchas the trajectorydistanceis still

simplisticin its presentform andratheradhoc.Futureworkusingmoresophisticatedtemporalmodelsto capturefacedynamicswith respectto subject,expression,movementandillumination changesis to beextensively conducted.

References

[1] V. Bruce,A. Burton,andP. Hancock.Comparisonsbetweenhumanandcomputerrecognitionof faces. In IEEE Inter-national Conferenceon AutomaticFace& Gesture Recog-nition, pages408–413,Nara,Japan,1998.

[2] V. Bruce,P. Hancock,andA.Burton.Humanfaceperceptionandidentification. In Wechsler, Philips,Bruce,Fogelman-Soulie,andHuang,editors,FaceRecognition: FromTheoryto Applications, pages51–72.Springer-Verlag,1998.

[3] T. Cootes,G. Edwards,andC. Taylor. Active appearancemodels. In EuropeanConferenceon ComputerVision, vol-ume2, pages484–498,Freiburg, Germany, 1998.

[4] T. Cootes,C. Taylor, D. Cooper, and J. Graham. Activeshapemodels- their trainingandapplication.ComputerVi-sionandImage Understanding, 61(1):38–59,1995.

[5] T. Cootes,K. Walker, and C. Taylor. View-basedactiveappearancemodels. In IEEE InternationalConferenceonAutomatic Face & Gesture Recognition, pages227–232,Grenoble,France,2000.

[6] G. Edwards,C. Taylor, andT. Cootes.Learningto identifyand track facesin sequences.In IEEE InternationalCon-ferenceon AutomaticFace & Gesture Recognition, pages260–267,Nara,Japan,1998.

[7] S. Gong, S. McKenna,and A. Psarrou. DynamicVision:From Images to Face Recognition. World Scientific Pub-lishing andImperialCollegePress,April 2000.

[8] S. Gong,A. Psarrou,I. Katsouli,andP. Palavouzis. Track-ing andrecognitionof facesequences.In EuropeanWork-shoponCombinedRealandSyntheticImage ProcessingforBroadcastand Video Production, pages96–112,Hamburg,Germany, 1994.

[9] B. Knight andA. Johnston.The role of movementin facerecognition.VisualCognition, 4:265–274,1997.

[10] Y. Li, S. Gong, and H. Liddell. Supportvector regres-sion andclassificationbasedmulti-view facedetectionandrecognition. In IEEE International Conferenceon Auto-maticFace& Gesture Recognition, pages300–305,Greno-ble,France,March2000.

[11] Y. Li, S. Gong, and H. Liddell. Constructingstruc-turesof facial identitiesusingKernelDiscriminantAnaly-sis. Technicalreport,QueenMary, University of London,2001. www.dcs.qmw.ac.uk/t yongmin/papers/kda.ps.gz,submittedto IJCV.

[12] Y. Li, S.Gong,andH. Liddell. Modellingfacesdynamicallyacrossviews andover time. In IEEE InternationalConfer-enceonComputerVision, Vancouver, Canada,July2001.

[13] B. Moghaddamand A. Pentland. Facerecognitionusingview-basedandmodulareigenspaces.In AutomaticSystemsfor the IdentificationandInspectionof Humans,SPIE, vol-ume2277,1994.

[14] S.Romdhani,S.Gong,andA. Psarrou.Onutilising templateandfeature-basedcorrespondencein multi-view appearancemodels. In EuropeanConferenceon ComputerVision, vol-ume1, pages799–813,Dublin, Ireland,June2000.

[15] S.Satoh.Comparativeevaluationof facesequencematchingfor content-basedvideoaccess.In IEEE InternationalCon-ferenceon AutomaticFace & Gesture Recognition, pages163–168,Grenoble,France,2000.

[16] L. Sirovich andM. Kirby. Low-dimensionalprocedureforthecharacterizationof humanfaces.Journal of Optical So-cietyof America, 4:519–524,1987.

[17] M. Turk andA. Pentland.Eigenfacesfor recognition.Jour-nal of CognitiveNeuroscience, 3(1):71–86,1991.

[18] R. Vanderbei. Loqo: An interior point codefor quadraticprogramming.Technicalreport,PrincetonUniversity, 1994.TechnicalReportSOR94-15.

[19] L. Wiskott, J. Fellous,N. Kruger, andC. Malsburg. Facerecognitionby elasticbunchgraphmatching. IEEE Trans-actions on Pattern Analysis and Machine Intelligence,19(7):775–779,1997.

[20] O. Yamaguchi,K. Fukui, andK. Maeda. Facerecognitionusingtemporalimagesequence.In IEEEInternationalCon-ferenceon AutomaticFace & Gesture Recognition, pages318–323,Nara,Japan,1998.