Dense Correspondence and Its Applications - Computer Vision for Visual Effects - page 189

t = 0 t = 0.25 t = 0.5 t = 0.75 t = 1

Figure 5.23. A real example of image morphing using dense correspondence fields. Top row:
intermediate images I 1 ; the original image I 1 corresponds to the leftmost image I 1 . Middle row:
intermediate images I 2 ; the original image I 2 corresponds to the rightmost image I 2 . Bottom row:
final morph sequence created by cross-dissolving between corresponding images in the top and
middle rows. The transformation is much more compelling than the simple cross-dissolve in
Figure 5.21 .

see that the cross-dissolve between corresponding images in the first and second
rows of Figure 5.23 yields a realistic morph between both the intensities and image
structures.
Morphing algorithms principally differ in their methods for obtaining the dense
correspondence fields between the image pair. Early methods used a compatible
quadrilateral mesh defined over the images (similar to the optimized-scale-and-
stretch grid in Section 3.5.1 ). Correspondences between points inside mesh quads
can be obtained by bilinear interpolation, or more generally using B-splines as dis-
cussed earlier in this chapter. The difficulty with this approach comes in attempting
to control the mesh to conform well to important image features, resulting in large
regions with either toomany or too fewmesh vertices. An alternate approach is to use
one of the scattered data interpolation techniques from Section 5.2 ; for example, Lee
et al. investigated both thin-plate splines [ 268 ] and adaptive, nonuniform B-spline
interpolating surfaces [ 269 ] to define the correspondences for morphing. A unique
aspect of the latter approach was the use of “snakes” [ 230 ] to automatically snap the
user-specified points to image features.
One of themost popular approaches to estimating dense correspondence formor-
phing is the field morphing technique proposed by Beier and Neely [ 35 ]. Unlike the
methods in Section 5.2 , the correspondence is interpolated from a set of several
corresponding user-drawn line segments on the two images. This allows the ani-
mator to have more control over the morph, since the method guarantees that the
correspondence between each pair of segments will be maintained in each mor-
phed image — something that a spline interpolation of feature matches cannot
guarantee. For example, to morph between two faces, an animator would draw
matching lines along the edges of the head, the eyebrows, the lips, and so on,
which is more intuitive than trying to establish feature matches in smooth, flat
regions.