A global-to-local model for the representation of human faces

In the context of face modeling and face recognition, statistical models are
widely used for the representation and modeling of surfaces. Most of these
models are obtained by computing Principal Components Analysis (PCA) on a set
of representative examples. These models represent novel faces poorly due to
their holistic nature (i.e.\ each component has global support), and they
suffer from overfitting when used for generalization from partial information.
In this work, we present a novel analysis method that breaks the objects up
into modes based on spatial frequency. The high-frequency modes are segmented
into regions with respect to specific features of the object. After computing
PCA on these segments individually, a hierarchy of global and local components
gradually decreasing in size of their support is combined into a linear
statistical model, hence the name, Global-to-Local model (G2L). We apply our
methodology to build a novel G2L model of 3D shapes of human heads. Both the
representation and the generalization capabilities of the models are evaluated
and compared in a standardized test, and it is demonstrated that the G2L model
performs better compared to traditional holistic PCA models. Furthermore, both
models are used to reconstruct the 3D shape of faces from a single photograph.
A novel adaptive fitting method is presented that estimates the model parameters
using a multi-resolution approach. The model is first fitted to contours
extracted from the image. In a second stage, the contours are kept fixed and
the remaining flexibility of the model is fitted to the input image. This makes
the method fast (30 sec on a standard PC), efficient, and accurate.