![](bg7.png)
established some application scenarios for EBGM.
EBGM seems a very suitable choice when there is
only one training image per class, and a reasonable
choice when there are two. Even for more training
images per class, EBGM should be used when the
images are extremely small, or when the eyes are not
ideally located.
A number of issues are still open. With regard to
identification accuracy, it is clear that we need to uti-
lize more appropriate metrics or enhance the existing
ones. One way to approach this problem is to weigh
the contribution of each fiducial feature by a differ-
ent amount when computing the total similarity over
the whole face. The major obstacle in this case is the
determination of the appropriate weights in a system-
atic way. A simple idea is to weigh each contribution
according to the expected accuracy in estimating the
feature position, so that we bias our decision towards
those fiducial points we have more confidence in.
Another way of improving performance would be
to use a larger number of fiducial points for each
image, by interpolating between the positions of the
known features. For example, 25 original points and
55 interpolated points have been used in (Bolme,
2003) to construct each Face Graph, while we have
used only the 20 points originally defined by Human-
Scan. The main concern here is to avoid using too
many and closely spaced points, as that would de-
grade identification accuracy (Wiskott et al., 1999).
Our early experiments show that adding just one ex-
tra point can improve average performance by about
1%, but some of the tried extra points can degrade
performance by three times as much.
The choice of kernel sizes is also very important as
images are downsized, since the use of smaller ker-
nels would reduce correlation between convolution
results from neighboring jets. Ideally, we would like
to have an algorithm that can dynamically adapt to the
image dimensions and adjust the size and composition
of the kernel set accordingly.
ACKNOWLEDGEMENTS
This work is partly sponsored by the EU under the
Integrated Project CHIL, contract number 506909.
REFERENCES
Belhumeur, P., Hespanha, J., and Kriegman, D. (1997).
Eigenfaces vs. fisherfaces: Recognition using class
specific linear projection. IEEE Trans. Pattern Analy-
sis and Machine Intelligence, 19(7):711–720.
Beveridge, J. and She, K. (2001). Fall 2001 update to CSU
PCA versus PCA+LDA comparison. Technical report,
Colorado State University, Fort Collins, CO.
Bolme, D. S. (2003). Elastic Bunch Graph Matching. Mas-
ter’s thesis, Computer Science Department, Colorado
State University.
Brunelli, R. and Poggio, T. (1993). Face recognition: Fea-
tures versus templates. IEEE Trans. Pattern Analysis
and Machine Intelligence, 15:1042–1052.
Christmas, B. and Galambos, C. (2005). http:
//www.ee.surrey.ac.uk/Research/
VSSP/RavlDoc/share/doc/RAVL/Auto/
Basic/Tree/Ravl.html.
Duda, R., Hart, P., and Stork, D. (2000). Pattern Classifica-
tion. Wiley-Interscience, New York.
Georghiades, A., Belhumeur, P., and Kriegman, D. (2001).
From few to many: Illumination cone models for face
recognition under variable lighting and pose. IEEE
Trans. Pattern Analysis and Machine Intelligence,
23(6):643–660.
Gunning, J. and Murphy, N. (1992). Neural network
based classification using automatically selected fea-
ture sets. IEE Conf. Series 359, pages 29–33.
Jesorsky, O., Kirchberg, K., and Frischholz, R. (2001).
Robust face detection using the Hausdorff distance.
In Audio and Video based Person Authentication -
AVBPA 2001, pages 90–95. Springer.
Martinez, A. and Kak, A. (2001). PCA versus LDA.
IEEE Trans. Pattern Analysis and Machine Intelli-
gence, 23(2):228–233.
Pentland, A., Moghaddam, B., and Starner, T. (1994).
View-based and modular eigenspaces for face recog-
nition. Proc. IEEE Conf. on Comp. Vision & Pattern
Rec., pages 84–91.
Pnevmatikakis, A. and Polymenakos, L. (2004). Compar-
ison of eigenface-based feature vectors under differ-
ent impairments. Int. Conf. Pattern Recognition 2004,
1:296–300.
Pnevmatikakis, A. and Polymenakos, L. (2005). A testing
methodology for face recognition algorithms. In 2nd
Joint Workshop on Multi-Modal Interaction and Re-
lated Machine Learning Algorithms, Edinburgh.
Turk, M. and Pentland, A. (1991). Eigenfaces for recogni-
tion. J. Cognitive Neuroscience, 3:71–86.
Waibel, A., Steusloff, H., and Stiefelhagen, R. (2004).
CHIL - Computers in the Human Interaction Loop.
In 5th International Workshop on Image Analysis for
Multimedia Interactive Services - WIAMIS 2004.
Wiskott, L., Fellous, J.-M., Krueger, N., and von der Mals-
burg, C. (1999). Face recognition by Elastic Bunch
Graph Matching. In Intelligent Biometric Techniques
in Fingerprint and Face Recognition, chapter 11,
pages 355–396. CRC Press.
Yang, M., Kriegman, D., and Ahuja, N. (2002). Detect-
ing faces in images: A survey. IEEE Trans. Pattern
Analysis and Machine Intelligence, 24(1):34–58.
Zhao, W., Chellappa, R., Rosenfeld, A., and Phillips, P.
(2000). Face recognition: A literature survey.
EBGM VS SUBSPACE PROJECTION FOR FACE RECOGNITION
137