FACE APPEARANCE SYNTHESIS USING FLEXIBLE APPEARANCE MODEL - 3D Face Processing: Modeling, Analysis and Synthesis - page 99

In Chapter 5, we have discussed how to generate face animations using geo-
metric MUs. To augment the synthesis with appearance variations, we propose
to decide the texture blending coefficients based on the corresponding geometric
MUPs. The intuition is that geometry and appearance are correlated such that
partial information about appearance can be inferred from the corresponding
geometry. In similar spirit, Zhang et al. [Zhang et al., 2003] have demon-
strated the effectiveness of synthesizing facial expression appearance details
from given geometry.
Compared to [Zhang et al., 2003], the geometry part of motion is fully de-
rived from our geometric-model-based synthesis. Thus we can design a simpler
formulation for the blending coefficients. In this scenario, the problem is to find
an appropriate texture given geometric shape: s . Suppose the corresponding ge-
ometric MUP is and the exemplars' geometric MUPs are
We define the blending the coefficient as
where is a constant which normalize the sum of blending coefficients to 1. In
practice, we need to avoid the blurring of the blending result. For this purpose,
we adjust the value of the constant experimentally such that there are only
N , (N < K) nonzero blending coefficients. Other coefficients are small and
can be set to zero.
4. Summary
We have described methods of face synthesis based on the flexible appearance
model. In particular, we have discussed two issues: (1) how to synthesize
illumination effects in face appearance; and (2) how to synthesize appearance
variations in face animations. The main contribution of this chapter is that we
show that our flexible appearance model can be used for synthesis in a flexible
way. More specifically, It means we can synthesize appearance variations based
on the model across different people and illumination environments.