5 Experiments

In this section, experiments using Gaussian mixture models of identity are described. Face image data were acquired and normalised in a fully automated way by the face tracking system. The neural network model used to perform tracking was trained using 9000 example face images rotated by

and scaled to

and

[ 8 ]. The normalised faces from the tracker therefore varied by at least these amounts in scale and rotation. Since the aim of these experiments was to compare methods for modelling identity rather than to optimise recognition accuracy, no attempt was made to reduce these variations.

Eight subjects were tracked through relatively unconstrained indoor scenes as they walked towards a fixed camera. Overhead lighting resulted in variations in facial illumination. The resolution of the area of the face tracked ranged from approximately

pixels when the subject was far from the camera to

pixels when the subject approached the camera. Two normalised face sequences were obtained for each subject. The first sequence of each subject was used for training and the second sequence for testing. In total, there were 326 training images and 296 test images. The number of training images per person varied from 21 to 60 and the number of test images from 21 to 53. Figure 5 shows 10 of the images used to form the training and test sets three of the people.

Face space was modelled by performing PCA on the training images. A specific model was computed from the training set. A generic model was computed using 644 of the images used to train a face detection neural network in the tracking system. These images were highly suitable, having similar variations in scale and rotation to the tracked data to be recognised. The training images were projected onto the first n ' eigenvectors and each person's identity was modelled by estimating either

with Gaussian mixtures. The 8 mixture models' parameters were stored along with the n ' eigenvectors and eigenvalues and subsequently used to perform classification of the test sequences.

Initially, both a specific and a generic eigenspace were computed using the first 40 eigenvectors. Table 1 shows a comparison of face classification using the specific and generic models. Identities were modelled by fitting a single radial Gaussian to each person's data. The percentage of images correctly classified for each person along with the percentage of total images classified correctly are given. Sequence classification results are also given based upon a majority vote i.e. the sequence is classified as the person with the most images. The result illustrates the fact that the use of a generic face space which could be used to facilitate identity verification, known/unknown or full recognition, in turn makes face classification more difficult.

Face Person (% images correct) Total Seq.

space
0 1 2 3 4 5 6 7 % (Maj.)

Specific 75 64 74 85 56 78 29 11 55.1 7

Generic 57 67 66 20 13 72 25 29 43.6 4

Table 1: Test set results with generic and specific face space models using 40 principal components. Identities were modelled by fitting a single radial Gaussian to each of the 8 people.

**Table 1:** *Test set results with generic and specific face space models using 40 principal components. Identities were modelled by fitting a single radial Gaussian to each of the 8 people.*
Face	Person (% images correct)	Total	Seq.
space	0	1	2	3	4	5	6	7	%	(Maj.)
Specific	75	64	74	85	56	78	29	11	55.1	7
Generic	57	67	66	20	13	72	25	29	43.6	4

Name M Tot. Seq.

type % Maj. Pr.

T-P 1 N 25.0 2 2

1-NN n N 32.1 1 1

T-P 1 Y 46.3 4 4

Radial 1 Y 44.3 4 4

Diag 1 Y 42.9 4 3

2-Rad 2 Y 52.0 5 7

3-Rad 3 Y 42.2 5 5

2-Diag 2 Y 41.9 4 5

Table 2: Test results with a 20-dimensional generic face eigenspace and identity mixture models. Column 2 indicates the number of Gaussians, M . Column 3 indicates the type of Gaussian where denotes a diagonal covariance, an independent variance and a variance equal to that of all other components. A `Y' in column 4 indicates that normalised pattern vectors were used.

**Table 2:** Test results with a 20-dimensional generic face eigenspace and identity mixture models. Column 2 indicates the number of Gaussians, M . Column 3 indicates the type of Gaussian where denotes a diagonal covariance, an independent variance and a variance equal to that of all other components. A `Y' in column 4 indicates that normalised pattern vectors were used.
Name	M			Tot.	Seq.
		type		%	Maj.	Pr.
T-P	1		N	25.0	2	2
1-NN	n		N	32.1	1	1
T-P	1		Y	46.3	4	4
Radial	1		Y	44.3	4	4
Diag	1		Y	42.9	4	3
2-Rad	2		Y	52.0	5	7
3-Rad	3		Y	42.2	5	5
2-Diag	2		Y	41.9	4	5

A reduction in the dimensionality of the generic face space from 40 to 20 did not result in any significant loss of accuracy. Face classification results using the 20-dimensional generic space are given in Table 2 . Sequences were classified (1) by a majority vote (Maj.) and (2) by accumulating probabilities (Pr.). Gaussian mixture models of various complexity were compared for modelling identity.

The first two methods in Table 2 used unnormalised pattern vectors. The first method (T-P) used single radial Gaussians of equal variance resulting in a nearest-mean classifier which was equivalent to the eigenfaces method of Turk and Pentland [ 10 ]. The second method was a nearest neighbour classifier (1-NN). Both these methods performed poorly. However, the use of normalised pattern vectors resulted in a significant improvement with T-P

classifying 4 sequences correctly. The mixture models had either radial or diagonal covariance Gaussians with between 1 and 3 components. A mixture of 2 radial Gaussians provided the best performance.