CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
80
difference [pix]
Figure 3: Two sample relative histograms of disparity differences
wrt reference disparity map. Note the scale difference between
both diagrams.
matching between FLI-MAP and Pictometry was tested. This is
interesting, because by this means the scene can be observed from
approximately the same direction through multiple views. The
overlap from consecutive Pictometry images is not large enough
to create 3-ray points, however, incorporating also FLI-MAP ima
ges makes this possible. Besides, this setup gives an interesting
geometry for forward intersection.
Two methods were used to assess the results: one quantitative
and one qualitative. For the quantitative assessment a reference
disparity map was computed from the FLI-MAP LIDAR data,
then the differences to the disparities from image matching were
analyzed using histograms. For a more qualitative assessment 3D
point clouds were computed from the matching results and then
assessed visually, also in comparison to the LIDAR point cloud.
Disparity map assessment For this assessment the reference
LIDAR points (density: 20 points per m 2 ) were projected into
the image plane as defined by the respective image orientation
and calibration parameters and subsequently a reference dispar
ity map was computed. Two issues are important here: first, only
first pulse LIDAR points should be considered, as also in the
image only the visible surface can be matched. Second, through
the oblique viewing direction as realized with the cameras one
has to take into account self-occlusion through buildings; the
laser scanner scans vertical and thus scans other parts of the scene,
especially on the backside of buildings visible in the images. To
avoid errors from that circumstance, only areas which do not
show these effects were used for the evaluation.
Figure 4: Results from dense matching in two overlapping South
looking Pictometry images. Top: left image and 3D cloud from
matching, center row: zoom to point cloud from matching at fa
çades (left) and top view (right), bottom row: point cloud color
coded height: reference (left), from matching (right)
thus matched points at façades which were not acquired by the
LIDAR device can not be assessed. Two of such relative his
tograms are shown in Fig. 3. The upper histogram shows the dif
ferences from the matching within two Pictometry images (see
Fig. 4). For this histogram approx. 50 • 10 3 matches were con
sidered (out of 2.2 • 10 6 in total), and around 70% of them show
a difference of ±3 pixels to the reference. The histogram at the
bottom shows the analysis from the matches within two oblique
images from FLI-MAP, refer to Fig. 5. For this histogram approx.
200 ■ 10 3 matches were considered (out of 6.4 • 10 6 in total).
Because of the smaller baseline between consecutive FLI-MAP
images, compared to Pictometry, the overlapping area is larger,
and thus results in more matches. Approximately 60% are within
the difference of ±3 pixels. All matches outside this tolerance
can be considered as blunder. A more in depth analysis revealed
that most blunders were caused in shadow areas or other areas
with poor texture. When assessing those histograms is should be
considered that errors from the image calibration and post esti
mation also contribute to those residuals, thus a final conclusion
on the absolute matching accuracy of the SGM implementation
can not be made.
Point clouds: Pictometry to Pictometry For the following
evaluations a forward intersection of the matched points was per
formed. A simple blunder detection was implemented by apply
ing a threshold to the residual for image observations. For two-
ray intersections this method can filter some blunders, but be-
The disparity maps were assessed by calculating the difference
disparity map and computing a histogram out of that one. Only
pixels showing a disparity value in both maps were considered,
