Graphics Reference
In-Depth Information
where the integers
m
,
n
0
specify the degree of themoment for each color channel.Mindruet al. showed various
ratios of sums and products of the color moments to be invariant to both geometric
(e.g., affine) and photometric transformations of an image, which can be concate-
nated to form a descriptor. Montesinos et al. [
333
] extended first-order grayscale
differential invariants to form a color descriptor.
The popularity of SIFT led to various attempts to “colorize” the SIFT descriptor.
An obvious approach is simply to concatenate the SIFT descriptors for a feature point
location computed in three color channels (e.g., RGB or HSV) to create a 3
≥
0 specify the order of themoment, and the integers
r
,
g
,
b
≥
128 =
384-dimensional descriptor, possibly reducing the dimensionality with PCA. van de
Weijer and Schmid [
509
] augmented the standard 128-dimensional SIFT descrip-
tor computed on luminance values with an additional 222 dimensions representing
color measurements, including histograms of weighted hue values and photometric
invariants. More straightforwardly, Abdel-Hakimand Farag [
1
] proposed to apply the
usual SIFT detector/descriptor to a color invariant image obtained as the ratio of two
linear functions of RGB. Both Burghouts and Geusebroek [
77
] and van de Sande et
al. [
508
] recently presented surveys and evaluations of color descriptors, generally
concluding that augmenting the SIFT descriptor with color information improved
repeatability results over using luminance alone.
Even if a grayscale feature detector/descriptor scheme is dictated by the appli-
cation, it may still be possible to improve performance by modifying the input to
the detector. For example, Gooch et al. [
174
] proposed an algorithm for processing
a color image into a one-channel image that differs from the traditional luminance
image. Instead, adjacent pixel differences in the new image are optimized to match a
function of color differences in the original image as well as possible. It may also be
possible to apply an algorithm like that of Collins et al. [
101
] to adaptively choose the
most discriminating one-dimensional space over the course of object tracking. For
example, the green channel may bemost discriminating for tracking a certain feature
on an actor in a natural environment, but a combination of red and blue channels
may be better as the actor crosses into a green-screen background.
×
4.5
ARTIFICIAL MARKERS
In a controlled environment such as a movie set, it's common to introduce
artificial
markers
for tracking — that is, patterns designed to be robustly, unambiguously
detected. Then, instead of using a generic feature detector, a customized detector
can be designed for rapidly locating and identifying such features. Themost common
use of artificial markers today is for augmented reality as opposed to visual effects.
Figure
4.22
illustrates several types of artificial markers, discussed inmore detail next.
Two-dimensional bar codes used to encode information are now commonplace,
for example in the context of tracking shipped packages. More recently,
QR codes
12
(Figure
4.22
a) have become popular for encoding information into a pattern of black
and white squares, which can be detected and decoded quickly by mobile devices
like cell phones. However, these patterns must fill a large percentage of an image to
12
QR Code is a registered trademark of Denso Wave Incorporated, Japan.
