In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
Figure 5: Results from dense matching in two overlapping FLI-
MAP images. Top: part of left image and point cloud from mat
ching, centre: 3D point cloud from matching, color coded height
(left) and color picked from images, bottom row: reference point
cloud and top view to matched point cloud . The circle indicates
an elevator box which is visible in the point cloud from matching,
but not in the laser scanning data.
cause only one redundant observation is available, quite a lot of
blunders will not be detected.
The point cloud as resulted from the triangulation of matches in
a Pictometry image pair are shown in Fig. 4. The top row in
that figure shows part of the left image and an overview on the
3D scene defined by the matched point cloud. The center row
shows a zoom in to the colored point cloud from matching, fo
cusing on some façades and the vertical view to that scene. Fi
nally, the bottom row shows the reference point cloud at the left
hand side, where the color codes the height (one full color cycle
equals 10m). The corresponding part of the matched point cloud
is depicted on the right hand side. From that figure some interes
ting observations can be made. Although most of the flat roofs
show poor texture, the corresponding part in the matched point
cloud is quite dense and homogeneous. However, the height of
the roof in the upper part is not correct, it is approx. 50cm lower
than shown in the reference point cloud. In the vertical view on
the point cloud from SGM the occluded areas are clearly visi
ble, whereas the vertical façades are not visible in the reference
point cloud. Overall, the structures are well represented, but the
mismatched pixels impose a certain level of noise to the scene.
Those mismatches can hardly be detected, as only stereo mat
ches are available in this case. The detailed zoom on the vertical
image shows that the accuracy in x-y direction is quite good, and
apparently even better than the estimated one (sx,y : 22 - 44cm).
Point clouds: FLI-MAP to FLI-MAP The triangulated point
cloud from the matching in two consecutive oblique images from
the FLI-MAP data is shown in Fig. 5. The zoom in to the col-
Figureô: Results from multiple view triangulation. Top: matched
point cloud, color and height, center: reference point cloud, bot
tom: top view
ored point cloud shows quite some details, for instance the eleva
tor box close to the statue is clearly visible. The statue itself is
well represented and even gaps in the matches where people were
walking during image exposure are clearly identifiable. However,
those views onto the point cloud were made from the viewing di
rection of the cameras, so the main error showing and effect in
viewing direction is not visible. The estimated depth accuracy (in
viewing direction, s' H ) of the along-track FLI-MAP data varies
from 90 to 130cm, and the error in X', L'-direction * s or *ly 2cm.
To assess the overall quality, the vertical view needs to be consid
ered: Here the uncertainty in viewing direction is quite obvious.
If the result from the vertical view-zoom is compared to the one
from the Pictometry data (Fig.4, center row), it can be observed
that the result from the FLI-MAP data is more inaccurate. This
visually achieved observation confirms the theoretically approxi
mated accuracy, which is about four times worse.
Point clouds: Multiple view For this experiment it was de
sired to exclude the wrong matches from the triangulation. To
achieve this goal, the dense matching results from several mat
ches were combined in the following manner: Dense matching
was performed in the three combinations: ® FLI-MAP 1 <-> FLI-
MAP2; © FLI-MAP2 <-> Pictometry 1; © FLI-MAP 1 *-* Pictom
etry 1. The matches from ® and © are linked in a way that the
matching points from the right image of ® are searched in the
left image of © and by this means corresponding matches ©’
FLI-MAP 1 Pictometry 1 are created. In the subsequent trian
gulation only those matches were considered which coincide with
the original matches from (D. Thus it can be expected that through
this double check some blunders were removed. For more de
tails on this method see (Gerke, 2008). In Fig. 6 some details on
the results are shown. The point cloud contains now less points
(1.6 • 10 6 points from the FLIMAP-only point cloud vs. 190 • 10 3
