Information Technology Reference
In-Depth Information
Local approaches
Local (or window-based) approaches are based on the similarity between two sets
of pixels. These methods are very popular for their simplicity and have been widely
used in computer vision for applications such as image registration, motion estima-
tion, video compression etc. In stereo correspondence, matching pixels consists in
comparing the neighborhood of the point for which a correspondence needs to be
established with the neighborhood of potential corresponding points located on the
associated epipolar line in the other image. Using a predefined search range, the
matching score for a pixel (
x
,
y
) at each allowed disparity
d
is derived by compar-
ing the intensity values of the window centered at (
x
,
y
) of the first view against the
window centered at the position (
x
+
d
,
y
) of the second image. Commonly used
matching measures include sum of absolute differences (SAD), sum-of-squared-
differences (SSD) and normalized cross-correlation (NCC). This latter measure is
insensitive to affine transformations between intensity images, which makes it ro-
bust to illumination inconsistencies that may occur between both views.
The choice of an appropriate window size and shape is crucial for window-based
local methods. Indeed, the use of windows of fixed size and shape may lead to er-
roneous matches in the most challenging image regions. In less-textured regions,
small windows do not capture enough intensity variations to make reliable match-
ing, whereas large windows tend to blur the depth boundaries and do not capture
well small details and thin objects. The different approaches for adaptive/ shiftable
windows [10, 11, 12] attempt to solve these problems by varying the size and shape
of the window according to the intensity variation. The work in [13] uses a multiple
window method where a number of distinct windows are tried and the one providing
the highest correlation is retained.
Progressive approaches
Progressive methods first establish correspondences between points that can be
matched unambiguously and then iteratively propagate the results of these matched
pixels on neighboring pixels [14, 15, 16]. The advantage of these approaches is their
computational cost, since they generally avoid a computationally expensive global
optimization. However, they propagate errors if an early matching point was not
well matched. The method of Wei and Quan [14] attempts to overcome this prob-
lem by matching regions derived from color segmentation. Since regions contain
richer information than individual pixels, the likelihood of early wrong matches is
reduced.
Cooperative approaches
Cooperative approaches make use of both local and global methods. They first cal-
culate a three dimensional space (
x
,
y
,
d
) where each element corresponds to the
pixel (
x
,
y
) in one reference image and all possible disparities
d
. A cooperative al-
gorithm is initialized by computing, for all possible matches, a matching score using
a local method. These initial matching scores are then refined by an iterative update
