AUGMENTED REALITY
Wearables
Perspective registration
Video Augmentation
The WristCam

RECENT PUBLICATION
Patents
Journals
Conferences

RESOURCES
Media Tech Resources

RETURN TO:
Userport homepage
York homepage

(1)	(2)	(3)	(4)	(5)	(6)

One approach to the finding and recognition of faces is to treat image windows as vectors of measurements (pel values), then apply conventional statistical pattern analysis on these inputs. For example, Principal Components Analysis applied to windows of pel values yields the "Eigenface" representation. Classification with classical methods and with neural networks has proved successful for both face/non-face discrimination and face recognition. However, reduced-dimensionality feature-spaces based just on pel values are not general-purpose enough to be used for all face processing tasks.

We have developed a face feature space derived via principal components from a set of measurements that includes deformation vectors and residues. The features set could also be augmented with edge strengths and frequency information. The largest principal components are used to form a classification subspace, and the animation above shows face variation along the first few of these components. Note that these top principal components have selected for certain kinds of variation in the training set: the first eigenvector tracks gross shape, the third, lighting direction, and the fifth, tilt. On the other hand, facial expression - specifically, smiling - is tracked by both the second and fourth eigenvectors (in the latter case, it is correlated with nose shape!) Using maximum likelihood classification within this space yields good performance for several face processing tasks, including recognition, pose estimation and expression analysis

Animations prepared by Qing Song

Q Song, J A Robinson, "A Feature Space for Face Image Processing", Proceedings of the International Conference on Pattern Recognition, Barcelona, September 2000, Volume 2, pp 97-100.
We propose criteria for a feature space for face image processing and a method for generating such a space. Beginning with many input dimensions, including deformation vectors (obtained through optical flow analysis between an input image and a neutral template) and deformation residues, we apply principal components analysis and Fisher's classification criterion to derive a feature space. We demonstrate classification in two important tasks -- face detection and expression analysis -- in each case using only one linear discriminant, thereby demonstrating that the feature space fulfils a restricted version of the criteria.