Graphics Reference
In-Depth Information
1
2 and a 2 =
Figure 2.5. Blue-screen matting using Equation ( 2.4 ) with a 1 =
1. We can see several
errors in the estimatedmattes, including in the interiors of foreground objects and the boundaries
of fine structures.
Vlahos [ 518 ] proposed many of the early heuristics for blue-screen matting; one
proposed solution was to set
α =
1
a 1
(
I b
a 2 I g
)
(2.4)
where I b and I g are the blue and green channels of the image normalized to the range
[
0, 1
]
, and a 1 and a 2 are user-specified tuning parameters. The resulting
α
values are
[
]
clipped to
0, 1
. The general idea is that when a pixel has muchmore blue than green,
α
should be close to 0 (e.g., a pure blue pixel is very likely to be background but a
pure white pixel isn't). However, this approach only works well for foreground pixels
with certain colors and doesn't have a strong mathematical basis. For example, we
can see in Figure 2.5 that applying Equation ( 2.4 ) results in amatte with several visual
artifacts that would need to be cleaned up by hand.
In general, when the background is known, Equation ( 2.2 ) corresponds to three
equations at each pixel (one for each color channel) in four unknowns (the fore-
ground color F and the
value). If we had at least one more consistent equation, we
could solve the equations for the unknowns exactly. Smith and Blinn [ 458 ] suggested
several special cases that correspond to further constraints — for example, that the
foreground is known to contain no blue or to be a shade of gray — and showed how
these special cases resulted in formulae for
α
α
similar to Equation ( 2.4 ). However, the
special cases are still fairly restrictive.
Blue-screen and green-screen matting are related to a common image processing
technique called background subtraction or change detection [ 379 ]. In the visual
effects world, the idea is called difference matting and is a common approach when
a blue or green screen is not practical or available. We first take a picture of the empty
background (sometimes known as a clean plate ) B , perhaps before a scene is filmed.
We then compare the clean plate to the composite image I given by Equation ( 2.2 ).
It seems reasonable that pixels of I whose color differs substantially from B can be
classified as parts of the foreground. Figure 2.6 shows an example inwhich pixels with
I
B greater than a threshold are labeled as foreground pixels with
α =
1. However,
Search WWH ::




Custom Search