....3....... 3... 3 4........ 7...... 0....0..........3.. 6... 6.... 7.... 8.....0........ 3.... 3.. 7.. 8.........30......33
........ 8.........9.....0 VRML...........4........5.... 7...... 8......... 9.... 9......... 0........................30 Cosmo VRML player....3..........5 MPEG-4.........6.....................
3
4
5
6 Roberto Bruell Tomaso Poggo Geometrc
7
8
9
40 80 MPEG-4 84 MPEG-4 <7.> <.5><.> VRML 3 MPEG-4 Feature Pots Extracto VRML 3 3 0
4 VRML,,, 3 3 3,, 3 3 3 j j j um um j j j / = = um: C m V V V 3 V m
* * w w V = = z y x V V V w * w * 3 3 3 3 3 3 3 V = = z y x V V V * * w w V = = z y x V V V * * 3 3 3 3 3 3 3 w w V + = = + + + z y x V V V * * w w Vm = = m z y m m x V V V
m S s = m k= Vk x Vjk x + Vk y Vjk y + Vk z Vjk z S.0~.0 W W = y = S SF : y S: S SF: S Global Face Local Face 6 S =.6 S P 5 S S N P : = P = P N 3
5 4
MDB SDB SDB SDB SA MDB MDB MA MW.0.9.9.0.9.0.9.3.0.9.6.9 Global Face.5.9 MDBBer 3D 5
. Uversty of Ber 9 58 MIT : 5.-5.4 Ber 9 MDB 5.5 MIT SDB MPEG-4 4 Ber 9 MDB : 9.8-9. 6.-6.4.-.4 7.,.6 MPEG-4 6
5. 3D Ber 9 3D 3D 90% 3 7 Global Face,45 3D 8 3D / % 4-74 9 54/58 93. 53-67 60-63 55 56/58 96.5-8,43-58 53 54/58 93. 68-770 45 5/58 89.6 8-336 55 4/58 7.4 0,9-3,34-37,74-77 Global Face 85 05 57/58 98. 0-8370 45 56/58 96.5 7
3D VRML A B A B A B A B A B 5.3 D D 3D 3D 4 9 D 0 8
D D 3D D 54/58 93.% 46/58 79.3%.9.0 56/58 96.5% 34/58 58.6%.5.9 54/58 93.% 35/58 60.3%.9. 5/58 89.6% 35/58 60.3%.9.0 4/58 7.4% 30/58 5.7%.0.9 57/58 98.% 44/58 75.8%.4.9 Global Face 56/58 96.5% 50/58 86.%.8.9 D VRML VRML A B C 9
D A B VRML A B A B A B A B S P 3D 96.5% D 87.9% 3 0
3D VRML A B D VRML A B
3D D 4/ 63.6% 3/ 59.0% 4/ 63.6% 7/ 3.8% 5/.7% / 9.0% 6/ 7.7% / 54.5% 7/ 77.% 6/ 7.7% 9/ 86.3% 4/ 63.6% Global Face 8/ 8.8% 8/ 8.8% Local Face 8/ 8.8% 5/ 68.% 3D 3D 3D D 68/80 85.0% 54/80 67.5%.0.9 69/80 86.% 39/80 48.7%.7.3 58/80 7.5% 36/80 45.0%.9.0 70/80 87.5% 49/80 6.%.9.3 58/80 7.5% 43/80 53.7%.0.9 75/80 93.7% 57/80 7.%.6.9 Global Face 74/80 9.5% 68/80 85.0%.9. Local Face 74/80 9.5% 66/80 8.5%
Precsso Rate = Recall Rate = 3D D 3D 3
: 4 3D D D 3D MPEG-4 <3.5> <3.6> <3.9><3.0> MPEG-4 eutral state <3.5><3.6> <3.9><3.0> : MPEG-4 3D D 6 D 3D : 8 MPEG-4 4
: 0 : MPEG-4 : D 3D <.5><.> 5
MPEG-4 :.. 6
7
[] Ferado Perera, MPEG-4 Facal Amato Techology: Survey, Implemetato, ad Results, IEEE Trasactos o Crcuts ad Systems for Vdeo Techology, Vol. 9, No., pp. 90-305, March 999. [] Sh ch Satoh et al., Name-It: Namg ad Detectg Faces News Vdeos, IEEE Multmeda, Vol. 6, No., pp. -35, Ja. Mar. 999. [3] Takeo Kaade, et al., Neural Network-Based Face Detecto, IEEE Trasactos o Patter Aalyss ad Mache Itellgece, Vol. 0, No., pp. 3-38, Jauary 998. [4] Jo Lous Betley, Multdmesoal Bary Search Trees Used for Assocatve Search, Comm. of the ACM, Vol. 8, No.9, pp.509-57, Sep. 975. [5] Ha Tao, et al., Compresso of MPEG-4 Facal Amato Parameters for Trasmsso of Talkg Heads, IEEE Trasactos o Crcuts ad Systems for Vdeo Techology, Vol. 9, No., pp.64-76, March 999. [6] Turk M. ad Petlad A., Face Recogto usg Egefaces, Proceedg of the IEEE, Computer Socety Coferece o Computer Vso ad Patter Recogto, pp.586-59, Mau, Hawa, Jue 99. [7] Hayua Wu, et al., Face Detecto From Color Images Usg a Fuzzy Patter Matchg Method, IEEE Trasactos o Patter Aalyss ad Mache Itellgece, Vol., No. 6, pp. 557-563, Jue 999. [8] Adreas Lats, et al., Automatc Iterpretato ad Codg of Face Images Usg Flexble Models, IEEE Trasactos o Patter Aalyss ad Mache Itellgece, Vol. 9, No. 7, pp. 743-756, July 997. [9] Ia Craw, et al., How Should We Represet Faces for Automatc Recogto? IEEE Trasactos o Patter Aalyss ad Mache Itellgece, Vol., No.8, pp. 75-736, August 999. [0] JA Goldste, el al., Idetfcato of Huma Faces, Proceedg of the IEEE, Vol. 59, No. 5, pp. 748-760, May 97. [] M. Vezjak, et al., Systems for Descrpto ad Idetfcato of Idvduals, Electrotechcal Coferece, 99 Proceedg, 6th Medterraea, Vol., pp. 5-54, 99. [] F. Goudal, et al., Face Recogto System Usg Local Autocorrelatos ad Multscale Itegrato, IEEE Trasactos o Patter Aalyss ad Mache Itellgece, Vol. 8, No. 0, pp. 04-08, October 996. 8
[3] F. Goudal, et al., Fast Face Recogto Method Usg Hgh Order Autocorrelatos, Proceedgs of Iteratoal Jot Coferece o Neural Networks, pp. 97-300, 993. [4] Sta. L ad Juwe Lu, Face Recogto Usg the Nearest Feature Le Method, IEEE Trasactos o Neural Network, Vol. 0, No., pp. 439-443, March 999. [5] aoguag Ja ad Mark S. Nxo, Extedg the Feature Vector for Automatc Face Recogto, IEEE Trasactos o Patter Aalyss ad Mache Itellgece, Vol. 7, No., pp. 67-76, December 996. [6] Roberto Bruell ad Tomaso Poggo, Face Recogto: Features versus Templates, IEEE Trasactos o Patter Aalyss ad Mache Itellgece, Vol. 5, No. 0, pp. 04-05, October 993. [7], MPEG-4 Processg of Workshop o the st Cetury Dgtal Lfe ad Iteret Techologes, 00. 9
30
3
7 Cosmo VRML player 3
33
34
35
36
37
38