the swath width of
-lines along strip,
essed. This size is
nedium and large-
n times and info
ferent sensors and
SING
irschmiiller (2005,
canner imagery is
| on the standardi-
the SGM core al-
ty post processing
| only minor adap-
bs. Therefore, the
putation and post
; under considera-
>s the theoretically
| by a number of
il disparity or, re-
dons every 45?) is
s can improve the
r current CPU im-
imum aggregated
image pixel leads
> roles of base and
pair are swapped,
tent disparities are
h occur predomi-
| Figure 3).
ion that small iso-
ight (or disparity)
s. For the required
ificantly different
s. Small segments
disparity map.
n optional cleanup
ights at the same
complex buildings
viewed off-nadir,
derived 2.5D pro-
ble 1), a reduction
carried out in two
lisparities into one
combination also
al accuracy; it re-
'd on local ranking
distribution of de-
t significant ones.
19%.
3.4 Info Cloud Generation
Both high-resolution and thinned disparity maps are projected
into object space, resulting in dense and thinned object point
representations, respectively. Multispectral information is as-
signed to each individual point in these info clouds. Classifi-
cation (water, low and high vegetation) based on the NDVI is
integrated into the approach and can be carried out if red and
NIR bands are available from the sensor. The final info cloud is
output into LAS format.
Figure 3: SGM processing workflow with disparity merge. Top:
ADS backward (left) and forward (right) panchromatic images.
Below: color-encoded disparity maps based on nadir/backward
(left) and nadir/forward image matching (right); stereo pair
merge in disparity space, with most occlusions and gaps filled.
Bottom: final RGB colored info cloud in true ortho view.
4. STEREO PAIR COMBINATION
It is normally required to combine SGM results from multiple
stereo views at some point in the workflow. Besides providing
results for large areas, this combination allows for gap filling
(occlusions) as well as consistency checks.
The ADS features three panchromatic view angles that are typi-
cally used in our processing, resulting in systematic and redun-
dant stereo coverage throughout the strip; the SGM results of
which can be merged at disparity level. Frame stereo depends
on the flight configuration, i.e. the image overlap along and
across strip. In any case, frame SGM jobs are inherently smaller
than the image size and, as opposed to the jobs along an ADS
strip, they overlap. The geometry of frame-based disparity maps
is dependent on the particular image and, therefore, not con-
sistent in a frame strip. As a consequence, frame-based SGM
results would have to be merged in object space.
4.1 Disparity Merge
The merge of the results from different stereo angles in disparity
space presumes common reference geometry, in case of the
ADS the epipolar rectified nadir view. Disparities from different
stereo pairs — nadir/backward and nadir/forward for the ADS —
are generally scaled relative to each other. For a line-scanner,
this scale can locally vary due to non-linear flight movement,
which we consider for the disparity conversion based on each
scan-line's well-known orientation. Scale-corrected disparities
are used to fill gaps that occur due to unavoidable occlusions in
individual stereo pairs as well as for consistency checks and
noise reduction by averaging.
The disparity merge for ADS stereo pairs is carried out before
any post processing steps, so that outlier elimination, cleanup
and thinning as described in section 2 are applied to the com-
bined disparity map. Figure 3 illustrates this merge, embedded
into the ADS workflow.
Figure 4: Merge of info clouds from individual frame stereo
pairs, based on DMC-II data. Top: section from a single strip,
flown West to East. Bottom left: corresponding info cloud from
the adjacent strip. Bottom right: Final result, combining data
from all three info clouds.
