Graphics Reference
In-Depth Information
Fig. 6.10
a
The proposed filter kernel, and
b
-
d
the outlier detection concept
If the area passes the consistency check in one of the dimensions, the depth pixel
z
v
—and therefore the joint image pixel
f
v
—is flagged as an outlier if
z
v
does not
exhibit the same consistency by exceeding a given threshold
˄
o
. Equation
6.7
shows
the outlier test when a depth consistency is noticed in the
X
-dimension, an analogous
test is used in case of consistency in the
Y
-dimension.
>˄
o
Z
v
(
x
−
ʻ,
y
)
+
Z
v
(
x
+
ʻ,
y
)
Z
v
(
x
,
y
)
−
(6.7)
2
After performing the proposed filter kernel, the patch centers are detected, as
conceptually represented in Fig.
6.10
b, c. Consistently, a standard morphological
grow algorithm is executed, which causes the detected center to grow only if the
neighboring pixels exhibit the same depth consistency as the initial outliers. As
depicted in Fig.
6.9
c, d, the complete patch is thereby detected. As a final step for
the patch filtering, the morphological grow is reversed and the detected patch is
filled with reliable depth values from its neighborhood. Since all of these opera-
tions are implemented on a pixel basis, they are inherently appropriate for imple-
mentation on a GPU, achieving a tremendous speedup compared to a generic CPU
algorithm.
6.3.3.2 Speckle Noise Filtering
Due to the nature of the human face, a significant amount of large homogeneous
texture regions are present. As indicated by [
24
], these areas cause the depth map to
contain spatial high frequency speckle noise. The noise is most effectively filtered
by a low-pass filter, but this eliminates the geometrical correctness of the depth map.
A standard 2D isotropic Gaussian filter is applied on the depth map and thanks to
its separable convolution properties, it can even be highly optimized on graphics
hardware [
11
].
