Information Technology Reference
In-Depth Information
We try to use landmark points that do not change positions with the change of fa-
cial expression. For example, the width of the eye always stays the same, but the
height changes depending on whether the eyelids are closed or not. Thus we try not to
use the landmark points that change positions easily.
Table 1.
Facial landmarks-dependent ROI boundaries. The notation
,,
represents the
x
and
y
coordinates from the
n-
th facial landmark.
Left Bound
Upper Bound
Height
Width
Left
Eyebrow
,
0.25
,
,
1.5
,
min
,
,
,
,
,
0.5
,
,
0.25
min
,
,
,
,
,
1.5
,
,
Right
Eyebrow
0.5
,
,
,
,
,
,
,
,
/4
0.5
,
,
1.9
,
,
1.4
,
,
Left Eye
,
0.6
,
,
,
0.5
,
,
,
,
/4
0.6
,
1.9
,
,
1.4
,
,
Right Eye
,
,
,
,
0.35
0.1
,
,
1.7
,
,
1.2
,
,
,
Nose
,
,
,
0.25
1.1
,
,
1.5
,
,
,
,
2
,
,
Mouth
1.1
.
,
,
,
,
3
1.5
,
,
6
,
,
,
4
,
,
Face
,
,
The bounding boxes used to crop out the ROI are shown in Table 4-1. They are
based on several ratios in the face that remain steady throughout the change of facial
expressions, such as the width of the eye or the height of the nose.
4.3
Feature Extraction
We use the histogram of the rotation-invariant LBP to describe each ROI. Since every
face is different, and to work around alignment issues, we use local regions. However,
within each region there is variation in position and shape. To solve the problem of
alignment in local scale, we introduce a sliding window in Fig. 5.
Search WWH ::
Custom Search