Graphics Reference
In-Depth Information
Figure 5.2
Real lens cameras exhibit distortion, which can be quantified from a photograph of a known
target (left). Once the distortion is quantified, it can be digitally corrected (right). (Courtesy
of Paul Debevec.)
5.1.2 Stereo Correspondence
Stereo correspondence is the process of finding corresponding points in images
captured from different viewpoints. This was once done manually by finding spe-
cific reference points in the photographs. An approach better suited to computer
processing is to attach special uniquely colored tags at strategic points on the
scene objects. The corresponding image pixels are then matched by color. How-
ever, this does require that the tags be placed on the objects before they are pho-
tographed. The general stereo correspondence problem assumes nothing about
the images, other than that they were captured from nearby viewpoints.
Suppose I 0 is a digitized image captured from one viewpoint, and I 1 is a sec-
ondary image captured from a nearby viewpoint. A pixel at
in the primary
image I 0 corresponds to a point on some scene object (actually, a small region
around the point on the object). The problem is to find the pixel in the secondary
image I 1 that corresponds to the same point on the same object. A naive approach
is to simply search the secondary image I 1 for a pixel having the same color as
the pixel at
(
x
,
y
)
in I 0 . The problem, though, is that pixel colors of corresponding
points may not be quite the same, due to imaging issues such as lighting variation,
noise,blurring, etc. Furthermore, the point imaged at the pixel
(
x
,
y
)
might be hid-
den by another object in the secondary image. A search based on a kind of local
averaging is used instead, which works under the assumption that the neighbor-
hood of pixels near
(
x
,
y
)
(
x
,
y
)
in I 0 corresponds to a pixel neighborhood in I 1 ,offset
by a vector
. The neighborhoods match best when the sum of the squares
of the pixel differences,
(
dx
,
dy
)
y
+
w y
x + w x
2
E
(
dx
,
dy
)=
w y (
I 1 (
k
+
dx
,
l
+
dy
)
I 0 (
k
,
l
))
,
(5.1)
k
=
x
w x
l
=
y
 
Search WWH ::




Custom Search