Face Detection and Modeling for Recognition - Biometrics Research ...

1.5.1 Face Alignment Using 2.5D Snakes
In our first recognition system (shown in Fig. 1.11), we have proposed a face modeling
method which adapts an existing generic face model (a priori knowledge of a human
face) to an individual’s facial measurements (i.e., range and color data).
We use
the face model that was created for facial animation by Waters [69] as our generic
face model. Waters’ model includes details of facial features that are crucial for face
recognition.
Our modeling process aligns the generic model onto extracted facial
features (regions), such as eyes, mouth, and face boundary, in a global-to-local way,
so that facial components that are crucial for recognition are fitted to the individual’s
facial geometry. Our global alignment is based on the detected locations of facial
components, while the local alignment utilizes two new techniques which we have
developed, displacement propagation and 2.5D active contours, to refine local facial
components and to smoothen the face model. Our goal of face modeling is to generate
a learned 3D model of an individual for verifying the presence of the individual in a
face database or in a video. The identification process involves (i) the modification
of the learned 3D model based on different head poses and illumination conditions
and (ii) the matching between 2D projections of the modified 3D model, whose facial
shape is integrated with facial texture, and sensed 2D facial appearance.
1.5.2 Model Compression
Requirements of easy manipulation, progressive transmission, effective visualization
and economical storage for 3D (face) models have resulted in the need for model
23