Information Technology Reference
In-Depth Information
training the system with five images per person from the first session and testing
with all the images from the second session i.e. 23 images of 83 subjects. Similar
results were achieved in both cases. Fig. 4-a shows the plot for the second case.
The recognition rate reaches its maximum with only 340 coecients.
We also studied the relationship of incident light and recognition accuracy and
found the recognition rates for individual images corresponding to one of the 23
illumination conditions. The system was trained with a single image per person
from the first session and then tested with a single image that corresponds to
the same illumination conditions from the second session. Results are reported in
Fig. 4-b. As expected, the images with frontal illumination yield high recognition
rates. Interestingly, for vertical stripe illumination, the recognition rate first
drops and then rises again as the stripe moves away from the center of the
screen indicating a non-linear relationship between the recognition accuracy and
lateral angle of incident light.
4.2 Experiment 2
In experiment 2, we study the relationship between the number of training im-
ages and recognition/verification rates. One or more images/person are used for
training and a single image/person is used for testing. We avoid testing all com-
binations of training images and take advantage or Lee et al.'s [13] findings that
one or two frontal and four to five laterally lit images are sucient for training.
This experiment was performed using the first session of our database, the
extended Yale B database and the CMU-PIE database. Fig. 5 shows the recog-
nition and verification rates, using our database, when 5 to 8 images per identity
are used for training and the remaining are used for testing. Table 1 summa-
rizes the results. Using 8 training images, the recognition and verification rate
at 0.001FAR was 99.87%. The 8 training images that gave the best performance
were number 2, 5, 12, 14, 17, 20, 21, 23. Note that this is consistent with the
findings in [13]. For fewer training images, we removed images that were lit from
large angles one by one in the following order 23, 21, 12, 20, 2, 17, 14 until we
100
100
99.9
99.9
99.8
99.8
99.7
99.7
99.6
99.6
99.5
99.5
99.4
99.4
99.3
99.3
99.2
8 training images
7 training images
6 training images
5 training images
99.2
8 training images
7 training images
6 training images
5 training images
99.1
99.1
99.0
99.0
1
3
5
7
9
11
13
15
17
19
10 −3
10 −2
10 −1
10 0
Rank
False Accept Rate (log scale)
(a)
(b)
Fig. 5. Experiment 2 results for our database. (a) CMC curves for 5 to 8 training
images and the corresponding (b) ROC curves.
 
Search WWH ::




Custom Search