Graphics Reference
In-Depth Information
We focused in this chapter on the basic mathematical formulations for natural
image matting. However, it's important to note that a key aspect of matting is real-
time user interactivity, and that much research has gone into semiautomatically
creating trimaps [
391
], quickly updating matte estimates based on additional scrib-
bles, and compositing the matte on its new desired background simultaneously with
estimation [
533
]. Several usability studies have been conducted on different methods
for segmentation and matting, such as [
280
].
Given the large number of matting algorithms proposed and the variety of
approaches to theproblem, it canbedifficult toknowhowtochoose a goodalgorithm.
Unfortunately, there is no singlebest algorithm, and the results of different algorithms
are difficult to numerically compare in a way that matches what a human thinks will
look “good.” For this purpose, Rhemannet al. [
392
] createda carefully ground-truthed
dataset for the impartial evaluation of matting algorithms, and maintain the website
alphamatting.com
with ranked results fromstate-of-the-art algorithms. This site is
an excellent resource for comparing, choosing, and developing matting algorithms.
They also proposed several new perceptually motivated error functions for quanti-
fying matting performance, noting that common metrics like mean-squared error
don't always correlate with visual quality.
Image segmentation is one of the major problems in computer vision; here, we
only discussed algorithms used as a starting point for matting. While graph-cut and
normalized-cut [
441
]- based methods have dominated the recent literature, other
approaches to interactive image segmentation include “intelligent scissors” [
339
]
and particle filtering [
363
].
In this chapter, we focused on extracting mattes from natural images, but in prac-
tice, the ultimate goal is to place thematte on a newbackground to create a composite
image. We'll discuss the process of combining elements from multiple images more
generally in Chapter
3
. Several topics discuss practical aspects of compositing for film
and video, notably Wright [
553
] and Brinkmann [
68
].
2.13
HOMEWORK PROBLEMS
Note: in the homework problems we often assume that image color channels and
intensities are in the range
[
]
[
]
0, 255
rather than
0, 1
for more direct compatibility
with image manipulation tools like GIMP and Matlab.
2.1 Take or find a photograph for which hard segmentation of the foreground
is likely to fail for high-quality compositing.
2.2
Suppose it was observed that the RGB values at pixel
i
were
I
i
=
]
. Determine
B
i
and
[
i
that are consistent with the hypotheses
that
F
i
= pure red, pure green, and pure blue, respectively. None of your
200, 100, 40
α
α
's
should be 0 or 1.
2.3
Suppose we obtain the clean plate
B
given in Figure
2.27
, and observe the
given image
I
. Determine twodifferent values for the images
F
and
that are
consistent with thematting equation: one that conforms to human intuition
and one that is mathematically correct but is perceptually unusual. Assume
that the intensity of the middle circle is 128.
α
