Page 414 -

Section 12.3 Registering Deformable Objects 382

Initial 3 it.s 8 it.s 11 it.s Original Converged

FIGURE 12.10: Active appearance models registered to face images. On the left, the initial
configuration of the model (blurry blob over the face; original face is second from right).
As the minimization process proceeds, the search improves the registration to produce, in
the final converged state, the registration on the right. Once we have this registration, the
location of the vertices of the mesh and the deformation parameters encode the shape of
the face. This figure was originally published as Figure 5 of “Active Appearance Models,”
by T. Cootes, G. Edwards, and C. Taylor, IEEE Transactions on Pattern Analysis and
Machine Intelligence, 2001, c IEEE, 2001.

Second, it is usually helpful to do all searches over scale. Using low-resolution
neutral and deformed images creates an objective function that changes less dra-
matically with changes of parameters, which makes the search easier; this gives a
good starting point for the search in a higher-resolution image. We could do this
by starting with low-resolution neutral and deformed images, estimating rotation
and translation, and then proceeding with increasingly high-resolution neutral and
deformed images, starting the search for rotation and translation estimates at the
point produced by the previous resolution. Once we have a rotation and translation
estimate, we estimate deformation starting at the lowest resolution and working up,
then polish rotation, translation, and deformation estimates starting at the lowest
resolution and working up. Finally, the best results seem to come from using quite
careful line searches (using either a gradient or the Newton direction).
The class of model we have described allows a rich range of variations. One
could filter or otherwise process the neutral and deformed images, thereby chang-
ing the objective function in important ways (for example, emphasizing high spatial
frequencies, or computing a vector of filter outputs to get a texture representation).
The method can be applied to 3D models as well, with the only major change be-
ing the increased complexity of 3D mesh topologies. Different deformation models
can be applied, and a wide range of search strategies have been used. Tim Cootes
publishes a variety of software tools for building, displaying, and using active ap-
pearance models at http://personalpages.manchester.ac.uk/staff/timothy.
f.cootes/software/am_tools_doc/index.html. There are also example datasets