Graphics Reference
In-Depth Information
S
seam
T
S
T
M
(a)
(b)
Figure 3.1. The compositing problem with hard-edged pieces. (a) Source, target, and mask
images. (b) In the composite image, regions from the source and target images are separated
by a seam. We want to make the transition between source and target as imperceptible as
possible.
for “twinning” effects, such as in Friends , Back to the Future II ,or Moon . For exam-
ple, an actor is filmed interacting with him/herself twice: once on the left side of the
screen and once on the right. In early versions of this effect, the seam between the
two shots was either very visible (e.g., a line down the middle of the screen) or hidden
by an obvious foreground object (e.g., a fence or tree). In this case, the problem is to
fuse two images I 1 and I 2 along a given seam, where neither image is naturally the
foreground or background and both are of equal importance.
Why were seams so visible using the simple technique of Equation ( 3.1 )? Even
if the camera was locked down with identical location and exposure to take both
the source and target shots, lighting conditions between shots are extremely dif-
ficult to match exactly, and the human visual system is extremely sensitive to
the presence of edges, especially in constant-intensity, low-frequency regions (see
Figure 3.3 a). The situation only becomes worse if one image is taken at a differ-
ent time or under different conditions than the other (for example, an actor shot
on a studio set is to be composited into an outdoor scene). Much of this chapter
is about the problem of hiding seams — both by choosing clever, non-straight-line
paths for the seams to take, and by more intelligently blending intensities across
the seam.
3.1.1
Weighted Transition Regions
An obvious approach to making the seam less noticeable is to blend the pixel
intensities across the seam with a parametric weighting function. That is, in the
neighborhood of the seam, the composite pixel color is a weighted average of the
source and target colors, depending on how close the pixel is to the source. For an
object that's not too complex, this weighted average looks like Figure 3.2 on a profile
perpendicular to the source boundary. This approach is similar to border matting
Search WWH ::




Custom Search