Facial Feature Analysis For Model Based Coding

Eric LarsonDecember 2007

Image Coding and Analysis Laboratory, Oklahoma State University

What is model-based coding? Facial Analysis Dealing with Dynamic Bandwidths

Solving a MOP quickly An application specific NSGA-II, with a

deterministic searchResultsConclusion

Alternative to sending raw video footage

Creation of “essential” parameters needed to reconstruct a scene

A real-time analysis nightmare

Copyright by Microsoft

Very Low Bit Rate Teleconferencing GamingMan-Machine InteractionVideo Telephony

Telephony for the deaf

Image Courtesy of Dr. Peter Eisert [3]

Analysis (by Synthesis)

Image Courtesy of Dr. Peter Eisert [3]

Images Courtesy of Dr. Peter Eisert [4]

Generously, Instituto Superior Technico

ISTface [22]

Gradient based approximation is not robust

Complication of direct optimization Handled by reducing FAPs

Do not address problem of dynamic bandwidth

Image Courtesy of J. Ahlberg [17]

Quality Objective Function:

FAP Number Objective Function:

D

2

10

255log 10 = PSNR

21

0

1

0

1),(

M

m

N

n

nmEMN

D

used is set fap if ,

used not is set fap if ,)( where

),(

1

0

18

1

iFAP

iFAPNi

FAP

Use NSGA-II for the multiple objective optimization

Assign a premature stopping criteriaChoose bandwidth Select FAP sets Use deterministic algorithm

Tournament selection used for crossover

Parents and children combined, sorted according to Domination Nearest Neighbor

RepeatFrom [7], NSGA-II

while {a search direction of improvement can be found} for {each dimension, step 20 units}

▪ -if the step is favorable, another step is made ▪ -Else, choose next dimension

find direction of steepest descent from original

point and improved point   while {step size scaling constant < 0.0001}

take step in the steepest descent direction▪ -if the new point is favorable, increase step size by two, ▪ -else, decrease step size by a factor of ten.

Update starting individual with new individual

Pareto fronts

Max Bandwidth (Uncompressed)

FRAME NO.

Selected FAP Setsa Best PSNRMean PSNR (Over 3 runs)

Mean Function Evaluations

Medium 0 0(3), 1(2), 2, 4, 5(2), 6, 9, 10, 11(2), 12, 13, 15(3), 16(3) 30.57 dB 30.36 dB 779

(~4.8 Kbits/s 1 0 (3), 1(2), 2, 4, 5 (3), 6 (2), 8, 9, 10, 11, 12, 13, 14(2), 15(3), 16(3), 17(3) 35.14 dB 32.54 dB 690

At 25 fps)b 2 0(2), 1, 2, 4, 5(2), 6(2), 7, 8, 10, 11(2), 12(2), 13(2), 14(3) , 15(3), 16(3), 17(2) 32.09 dB 29.50 dB 392

3 0(2), 1, 2(2), 5, 6(2), 7(2), 8, 9(2), 11(2), 12, 13(2), 14(2), 15(3), 16(3), 17 33.20 dB 29.99 dB 5614 0(2), 1, 2(2), 3, 5(2), 6(2), 7, 8, 10(2), 11(2), 13(2), 14(2), 15(3), 16(2), 17(2) 32.98 dB 28.14 dB 415

5 0(2), 1(2), 2(2), 3, 6(2), 7(2), 8, 9(2), 10(2), 11, 12(3), 13, 14(3), 15(3), 16(2), 17(2) 32.90 dB 28.73 dB 299

6 0(2), 1(3), 2, 5, 7(2), 8(3), 9, 10(2), 11, 12, 14, 15(3), 16(3), 17 32.13 dB 30.89 dB 7487 0(3), 1(2), 4, 5, 6, 7(3), 8(3), 11(2), 12(2), 13, 14, 15(3), 16(3), 17(2) 31.91 dB 29.51 dB 4458 0(3), 2, 4, 5(2), 6(2), 8, 9, 11(2), 12(2), 13(2), 14(2), 15(3), 16(3), 17(2) 30.97 dB 29.53 dB 7269 0(3), 1(2), 3, 5(2), 6(2), 7, 8, 9(2), 10(2), 11(2), 12(2), 14, 15(3), 16(3), 17 30.96 dB 28.99 dB 451

10 0(3), 2, 3, 5, 6, 7, 8, 9, 10(2), 11(2), 12(2), 13(2), 14(2), 15(3), 16(3), 17(2) 30.21 dB 28.80 dB 527

Low 0 0, 7, 8(2), 11(2), 14(2), 15(3), 16(2) 29.95 dB 27.13 dB 573(~2.4 Kbits/s 1 0, 5, 8, 11(2), 12, 14, 15(3), 16(2), 17(3) 33.23 dB 29.46 dB 595At 25 fps)b 2 8, 10, 11, 12(2), 13(2), 14, 15(3), 16(2), 17(3) 32.02 dB 27.21 dB 773

3 2, 5, 6, 8, 9, 12(2), 14, 15(3), 16, 17 28.77 dB 24.34 dB 8084 1, 9(2), 10, 11, 12(2), 14(2), 15(3), 17(3) 22.99 dB 22.80 dB 7455 1, 2, 4, 5, 6, 9, 11, 12, 14, 15(3), 16(2), 17 29.25 dB 26.93 dB 4466 2, 5, 6, 9(2), 10, 11(2), 12, 14(2), 15(2), 16(3), 17 29.67 dB 25.75 dB 376

7 1, 2, 7, 8, 9, 10, 12, 14, 15(3), 16(3), 17 29.01 dB 28.41 dB 3868 1, 3, 9, 12, 13, 15, 16(3) 28.97 dB 23.98 dB 5299 0, 5, 9, 10, 11, 12, 15(2), 16(3), 17 28.79 dB 25.93 dB 694

10 3, 5(2), 6(2), 9, 10(2), 12, 15, 16(3) 27.56 dB 24.25 dB 226

Histogram of all resultant individuals

Video Sequence

Frame 90

Low

Medium

Frame 93

Low

Medium

Frame 96

Low

Medium

Frame 99

Low

Medium

Frame 102

Low

Medium

Frame 105

Low

Medium

Frame 108

Low

Medium

Frame 111

Low

Medium

Frame 114

Low

Medium

Frame 117

Low

Medium

Frame 120

Low

Medium

Deficiencies can be traced back to selection of PSNR

Future work should include error functions like SSIM or Eigen-faces

Algorithm works Accentuates the details of PSNR

1. D. Pearson, “Developments in model-based image coding,” Proceedings of the IEEE, Vol. 83, No. 6, June 1995.2. I. Pandizic. J. Ahlberg, M. Wzorek, P. Rudol, and M. Mosmondor, “Faces Everywhere: Towards Ubiquitous Production and

Delivery of Face Animation,” Proceedings of the 2nd international conferenice on mobile and ubiquitous media, 20033. P. Eisert, “MPEG-4 facial animation in video analysis and synthesis,” International Journal of Imaging Systems and

Technology, June 2003.4. P. Eisert, “Very Low Bit Rate Coding,” Doctoral Thesis, November 2000.5. J. D. Schaffer, “Multiple objective optimization with vector evaluated genetic algorithms,” 1st international conference on

genetic algorithms, 1985.6. K. Deb, “Multi-objective genetic algorithms: problems, difficulties, and construction of test problems,” Evolutionary

Computation, 1999.7. Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T., A fast and elitist multiobjective genetic algorithm: NSGA-II, IEEE

Transactions on Evolutionary Computation, 2002.8. F. I. Parke, Parameterized Models for Facial Animation, IEEE Transactions on Computer Graphics and Animation, 1982.9. R. Forchheimer and T. Kronander, “Image coding – from waveforms to animation,” IEEE Transactions on Acoustics, Speech,

and Signal Processing, 37:1212, 1989.10. C. S. Choi, K. Aizawa, H. Harashima, and T. Takebe, “Analysis and synthesis of facial image sequences in model-based

image coding,” IEEE Transactions on Circuits and Systems for Video Technology, June 1994.11. M. Buck, “Model based image sequence coding,” Motion Analysis and Image Sequence Coding, Ch. 10, Kluwer Academic

Publishing, 1993, pp. 285-315.12. N. Diehl, “Object motion estimation and segmentation on image sequences,” Signal Processing: Image Communications,

Vol. 3, No. 1, February 1991, pp. 23-56.13. K. Aizawa, H. Harashima and T. Saito, “Model-based analysis-synthesis image coding (MBASIC) system for a person’s face,”

Signal Processing: Image Communication, vol. 1, pp. 139-152, 1989.14. I. S. Pandizic and R. Forchheimer, “MPEG-4 Facial Animation: the Standard, Implementation, and Applications,” 1st Ed. John

Wiley and Sons, 2002, pp. 3-41.15. J. Ahlberg and R. Forchheimer, “Face Tracking for model-based coding and face animation,” International Journal on

Imaging Systems Technology, Wiley Periodicals, Vol. 13, pp. 8-22, 2003.16. Dornaika, F., Ahlberg, J., Fast and Reliable Active Appearance Model Search for 3D Face Tracking, Proceedings of Mirage

2003, March 2003.17. Dornaika, F., Ahlberg, J., Fitting 3D Face Models for Tracking and Active Appearance Model Training, Image and Vision

Computing 24(2006), Science Direct, 2006.18. Carter, E.F, 1994, The Generation and Application of Random Numbers, Forth Dimensions, Vol XVI, Nos 1 & 2, Forth Interest

Group, Oakland California.19. S. Kirkpatrick, C. D. Gelati, and M. P. Vecchi, “Optimization by simulated annealing,” Science, Vol. 220, No. 4598, pp. 671-

680, 1983.20. T. Edgar, D. Himmelblau, and Lasdon, L., Optimization of Chemical Processes, 2nd Edition, McGraw-Hill, New York, NY, 2001.21. G. Reklaitis, A. Ravindran, and Ragsdell, K., Engineering Optimization, Methods and Applications, 2nd Edition, John Wiley and

Sons, New York, NY, 2006.22. ISTface, Program from Instituto Superior Technico, standard FAP animation sequence, “wow25.fap”.23. J. Jiang, A. Alwan, P. A. Keating, and T. A. Edward Jr., “On the relationship between face movements, tongue movements,

and speech acoustics,” EURASIP Journal on Applied Signal Processing, 2002.24. Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Image Quality Assessment: From Error Visibility to Structural Similarity,”

IEEE Trans. Image Process. 13, 600–612 (2004).