Information Technology Reference
In-Depth Information
100
80
60
40
20
TST
TRN
Hausdorff+MLP
0
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
x
Fig. 10.11.
Localization performance: percentage of examples having small
d
eye
for the pro-
posed method (TRN, TST) and for the hybrid system (Hausdorff+MLP) [111].
in comparison to the data taken from [111] (Hausdorff+MLP). All training examples
have been localized successfully. The performance on the test set is also good. Only
1.5% of the test examples have not been localized accurately enough. Compare this
to the 8.2% mislocalizations in the reference system.
A detailed analysis of the network's output for the mislocalizations showed that
in these cases the output is likely to deviate from the one-blob-per-eye pattern. It
can happen that no blob or that several blobs are present for an eye.
By comparing the activity
a
blob
of a segmented blob to a threshold
a
min
= 3
and
to the total activity of its feature array
a
total
, a confidence measure
c
is computed
for each eye:
=
a
blob
/a
total
,
c
1
1
:
a
blob
> a
min
=
c
2
,
:
else
a
blob
/a
min
c
=
c
1
ยท
c
2
.
(10.3)
The confidences of both eyes are multiplied to produce a single localization confi-
dence. Since the faces in the BioID database are mainly in an upright position, the
confidence is reduced if the blobs have a large vertical distance, compared to their
horizontal distance. Figure 10.8 shows some example confidences. Figure 10.12
displays the confidence versus the relative eye distance
d
eye
. One can observe that
high distances occur only for examples with low confidences. Furthermore, exam-
ples with high confidence values have low distances. Thus, the confidence can be
used to reject ambiguous examples.
The localization confidence is compared to a reject threshold. In Figure 10.13,
one can see that rejecting the least confident test examples lowers the number of
Search WWH ::
Custom Search