Graphics Reference
In-Depth Information
this, then either it always looks flat or it hurts in some places, and you don't get a
sense of stereo throughout the shot.
RJR: Do you ever need to touch up the disparity field after it's been acquired by the
stereo rig?
Beier: We only do that to fix problems, like if the rigs were badly misaligned. We
would never take a live-action element and move it further back. In a few cases, we'll
move a CG element to line up with an actor's eyeline so he seems to be looking at
it. It would be nice to be able to move stereo elements around after the fact — for
example, to bring people in the background forward a bit so they don't all seem to
be the same depth — but that's really hard to do since it requires stereo rotoscoping
and inpainting. That's one advantage of 3D conversion.
5.10
NOTES AND EXTENSIONS
We generally assumed in this chapter that the dense correspondence field between
a pair of images reflects an underlying physical reality — that is, that each pair of
corresponding points arises from some single point in the scene. However, dense
correspondence doesn't have to be physically meaningful to be useful. For example,
in some view synthesis applications, all that reallymatters is whether the synthesized
image is plausible, not whether it's physically consistent with the underlying scene.
This is especially true in applications like video coding, where we often just want a
good prediction of what an intermediate image will look like.
We focusedondense correspondence ina generic sense,meaning thatwe assumed
noknowledge about the contents of the images.Whenwe knowthe images come from
a certain class, then it would be advisable to use class-specific detection algorithms
to obtain better correspondence. For example, if we know the images are close-
up views of faces (e.g., for a morphing application), we could apply a customized
active appearance model (e.g., [ 180 ]) to immediately obtain a meaningful dense
correspondence map between them, matching eyes, noses, mouths, and so on.
Outside of visual effects, the medical imaging community is extremely interested
in algorithms for deformable image registration , which can be viewed as a type of
optical flow problem. Generally, the goal is to warp one image to the coordinate sys-
tem of a second, for example, to compare disease progression in images of of the
same patient over time, or to compare similar images of different patients to create
an “atlas.” The algorithm proposed by Joshi and Miller [ 228 ] in Section 5.2.3 is one
example of this application. Holden [ 201 ] gives a reviewof deformable image registra-
tion techniques for medical image analysis. Medical image registration is often posed
using the framework of fluidflow; that is, the image pixels are treated as a viscous fluid
that deforms subject to the rules of continuum mechanics. The deformation field is
usually obtained by solving a partial differential equation, and the process may be
very slow to converge (see, e.g., [ 94 ]). A popular method that resembles an iterative,
multiscale optical flow algorithm was proposed by Thirion [ 488 ]. However, like the
methods in Section 5.2 , methods for medical image registration are not designed to
handle occlusions or discontinuities in the flow fields, and could generally benefit
Search WWH ::




Custom Search