Graphics Reference
In-Depth Information
Again, there are some academic matting algorithms that work wonders on still
photographs, but the minute you have a video with wispy hair blowing in front of
a forest, you can get a solution but it's not natural. It's critical that the alphas have
that spatial softness and temporal continuity that you need for believability — you
can't have the big thick line around Godzilla that everyone tolerated in the '60s! A
third-party vendor may come out with a software plug-in for natural image matting,
but unless it really nails every problem you throw at it, it's not going to become a
standard.
2.12
NOTES AND EXTENSIONS
Matting for Hollywood movies was pioneered by Petro Vlahos and his son, Paul Vla-
hos, who patentedmany techniques related to blue-screen and green-screenmatting
and compositing from the late 1960s to the early 2000s. They won several Oscars
for their contributions and founded the Ultimatte Corporation, which produces an
industry-standard product for blue- and green-screen matte extraction.
Originally, mattes were strips of monochrome film that were exposed (i.e.,
transparent) in regions corresponding to the foreground and opaque (i.e., black)
elsewhere — an analogue of the alpha channel discussed in this chapter. For early
special effects, different elements and mattes were laboriously segmented, aligned
and sandwiched together to produce a final composite for such films as
Mary Pop-
pins
,
Superman
, and the original
Clash of the Titans
. For more on the early history of
matting in film, see the topic by Rickitt [
393
].
While Chuang et al. were the first to put the matting problem in the Bayesian
context, Ruzon and Tomasi [
412
] had previously proposed a related algorithm that
involved mixtures of isotropic Gaussian distributions (i.e., diagonal
i
) to model the
foreground and background. This is viewed as one of the first principled natural
matting algorithms from the computer vision/graphics community.
Singaraju et al. [
454
] showed that when the foreground or background intensities
of a window around a pixel are less general than the color line assumption (i.e.,
either or both is of constant color), then the closed-form matting equations permit
more degrees of freedom than necessary. They showed how to analyze the rank of
image patches to create an adaptivematting Laplacian that outperforms closed-form
matting in these situations.
We note that while most of the algorithms described here represented colors
using the usual RGB values, and measured similarity using Euclidean distance in
this space, some authors have recommended using a different color space that
better reflects when two colors are perceptually similar. For example, Ruzon and
Tomasi [
412
], Bai and Sapiro [
25
], and others proposed to use the CIE Lab color
space for matting operations, while Grady [
178
] used the linear transform of RGB
defined by the locality-preserving projections algorithm [
193
]. For Poisson matting,
Sun et al. [
478
] recommended a linear transform of RGB that is computed to mini-
mize the variance between the background samples. Another possibility would be to
use a higher-dimensional set of filter responses (e.g., a Gabor filter bank applied to
luminance as in [
377
]) which might be more sensitive to local texture in an image.
