Face Localization - Hierarchical Neural Networks for Image Interpretation - page 209

0.5 3

TRN TST

0.45

2.5

0.4

0.35

2

0.3

0.25 1.5

0.2

1

0.15

0.1

0.5

0.05

0 0

0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

(a) (b)

confidence confidence

Fig. 10.12. Face localization performance. The confidences are shown versus the relative eye
distance d eye for: (a) the training set; (b) the test set.

1.6

TST

1.4

1.2

1

0.8

0.6

0.4

0.2

0

0 1 2 3 4 5 6 7 8

reject in %

Fig. 10.13. Localization performance: rejecting the least confident examples lowers the num-
ber of mislocalizations.

c = .000 c = .000 c = .004 c = .017
Fig. 10.14. Test examples with lowest confidences. The highest resolution network inputs,
the outputs, and the confidences are shown.

mislocalizations rapidly. When rejecting 3.1% of the images, only one mislocaliza-
tion is left. The average localization error of the accepted examples is d eye = 0 . 06 .
That is well within the area of the iris and corresponds to the accuracy of the given
eye coordinates.
Figure 10.14 shows the four test examples with the lowest confidences. In the
leftmost example, the network has produced an extra blob for the mouth which pre-
vents segmentation of the blob that corresponds to the eye. Since the person's face