Page 412 -
P. 412

Section 12.3  Registering Deformable Objects  380



















                              Reference points           Relaxed points        Relaxed intensity

                            FIGURE 12.8: A set of reference points placed over a face, on the left.At the center,
                            these points in a relaxed configuration. Now assume we have a reasonable triangulation
                            of the original set of points. By placing those points in correspondence with the relaxed
                            configuration, we can map the intensities of the reference face to a relaxed configuration
                            (right). This figure was originally published as Figure 1 of “Active Appearance Models,”
                            by T. Cootes, G. Edwards, and C. Taylor, IEEE Transactions on Pattern Analysis and
                            Machine Intelligence, 2001, c   IEEE, 2001.


                            (i.e., that the (s, t) values naturally interpolate between the vertices of the triangle).
                            We can then produce a neutral image of the face simply by moving the vertices to
                            their neutral position (Figure 12.8).
                                 There is nothing special about the neutral locations of the mesh vertices; we
                            can generate an intensity field for any configuration of these vertices where triangles
                            don’t overlap. This means we can search for the location of a deformed triangle
                            in a new image I d by sampling (s, t) space at a set of points (s j ,t j ), and then
                            minimizing
                                                                                2
                                                g(||I d (p(s j ,t j ; w)) −I n (p(s j ,t j ; v))|| )
                                              j
                            as a function of the vertices w i . Here, as before, if we do not expect outliers, then
                            g is the identity, and if we do, it could be some M-estimator. If we expect that the
                            illumination might change, then it makes sense to minimize
                                                                                   2
                                              g(||aI d (p(s j ,t j ; w)) + b −I n (p(s j ,t j ; v))|| )
                                            j

                            as a function of the vertices w i and of a, b.
                                 When there is more than one triangle, the notation gets slightly more com-
                            plicated. We write v (k)  and w (k)  for the vertices of the kth neutral and deformed
                            triangles respectively. We do not expect the vertices to move independently. A vari-
                            ety of models are possible, but it is natural to try and make the model linear in some
                            set of parameters. One reasonable model is obtained by writing V =[v 1 ,... , v n ]
                            (resp. W =[w 1 ,..., w n ]) for the 2 × n matrices whose columns are the vertices of
                            the neutral (resp. deformed) points. Now we have a set of r 2×n basis matrices B l ,
   407   408   409   410   411   412   413   414   415   416   417