Graphics Reference
In-Depth Information
To extend a trained appearance-based exemplar model to new conditions, an
online EM-based algorithm is used to update the appearance model of these
exemplars progressively.
1.1 Feature extraction
We assume faces are Lambertian. Let denote the neutral
face texture and deformed face texture, respectively. We denote
As pointed out by Section 2.2.1 of Chapter 6, the ratio image is inde-
pendent of surface reflectance property [Liu et al., 2001a]. Therefore,
can be used to characterize facial motions of faces with different albe-
To use in face tracking, more compact features need to be extracted
from the high dimensional ratio image. Because low frequency variation of
facial motion could be captured by geometric-based methods, we extract fea-
tures from in frequency domain and use the high frequency components
as the features for motions not explained by geometric features. Second, past
studies on facial motions [Zhang et al., 1998, Tian et al., 2002] have shown
that there are certain facial areas where high frequency appearance changes are
more likely to occur and thus suitable for texture feature extraction. We apply
this domain knowledge in our feature extraction. However, because of noise in
tracking and individual variation, it is difficult to locate these locations automat-
ically with enough precision. Therefore, we extract the texture-based features
in facial regions instead of points, and then use the weighted average as the final
feature. Eleven regions are defined on the geometric-motion-free texture map.
These eleven regions are highlighted on the texture map in Fig. 7.3. Note that
these regions can be considered constant in the automatically extracted texture
map, where the facial feature points are aligned by geometric tracking.
Figure 7.3. Selected facial regions for feature extraction.
Search WWH ::

Custom Search