a-sets.
Image
scale
1:4000
1:2500
1:6000
1:2500
1:4000
1:4000
o airplanes or
|.
| truth") and
ere measured
ses, the loca-
ber locations.
ured in order
ig results. In
ber of points
five columns
,5 and 6 pho-
lata-sets were
ig the original
PE sets, only
rements were
The available
| WY sets was
-PE-30yum.
were checked
|| points were
to distinguish
parameters of
ication of the
em of the first
system to be
results of the
vo columns of
n residual for
| TEXAS and
e of the other
"this. First,
z them on the
measure the
e by creating
Table 3: Parameter setting for the iterative procedure.
Iteration 1 2 3
Pixel size [um] 240/180 120/90 60/45
Window size 13 x 13 23 x 23 | 45 x 45
# of matched points 5 x5 9ix132|:174
Grid interval [um] 1920/1440 | 960/720 | 480/360
Table 4: Results of the matching procedure.
00 rsd. # of ii rjctd
OSU 5 6 ^ 96 12
SWISS-2 6 9 96 12
wy 6 8 96 0
SWISS-3 6 11 166 18
TEXAS 4 13 376 22
OEEPE 12 46 734 102
“models” with images from different strips. However, meas-
uring exactly the same point on more than two images is not
possible. The second reason is the relatively low resolution
levels (22.5 um and 30 um, respectively) that were available
for the images of these sets. Nevertheless, an overall accuracy
of one fifth of a pixel was obtained for the TEXAS set, and
approximately one third of a pixel for OEEPE.
4.2 Results
The manually measured reference points provided approx-
imations for the matching procedure. Normally, approxima-
tions that are obtained through an automated procedure (like
AATS) are not as good as those considered here. However,
preliminary experiments (see Krupnik (1994)) showed that
adding errors of up to 2 pixels to the coordinates obtained
from the reference data hardly changed the results. Fur-
thermore, the main idea of the object-space approach is to
improve the accuracy of the results by improving the math-
ematical model of the matching, assuming that the reliability
problem has been solved by increasing the size of the match-
ing window. Therefore, it was decided not to add errors to
the reference image coordinates.
For the iterative procedure, three levels were chosen. The
pixel size, the size of the matching windows, the number of
matched points and the grid interval for each iteration are
shown in Table 3. In the entries of the table that contain two
values, the first value refers to the OSU, WY, SWISS and
OEEPE data-sets, and the second value refer to the TEXAS
data-set. As can by observed, the resolution of the final level is
lower than the available resolution. This enabled a comparison
with the results obtained by the manual measurements.
Table 4 shows the results of the matching in the object space.
For each data-set, the overall accuracy, the maximum residual,
the total number of observations and the number of rejected
observations are shown. It should be noted that in order to
simplify the testing procedure, for each point where a blunder
was detected (i.e., the matching diverged or converged to a
wrong location), all the observations were not considered.
In the case of only two images, the results can be easily com-
pared to results of a traditional matching procedure. A stand-
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Table 5: Comparison between object-space and image-space
approaches.
Object-space Image-space
00 rsd. | rjct. Jo rsd. | rjct.
OSU 5 6 12 7 13 20
SWISS-2 6 9 12 9 11 16
WY 5 8 0 5 6 12
Table 6: Comparison between the matching results and ref-
erence measurements.
Matching results Ref. measur.
To rsd. To rsd.
OSU 5 6 2 2
SWISS-2 6 9 3 4
WY 6 8 5 3
SWISS-3 6 11 4 7
TEXAS 4 14 5 12
OEEPE 12 40 11 31
ard least-square matching was applied to the OSU, WY and
SWISS-2 data-sets. The same approximations and window
size (45 pixels) as in the object-space case were used. The
results are compared in Table 5. Since not the same points
were rejected in both cases, only points that were considered
in both cases were used in the comparison.
The comparison shows a clear advantage of the object-space
approach on the image space approach for the first two data
sets. Both oo and the maximum residual are smaller in the
results of the object space approach. For the WY data set,
there is no significant difference between the results of the two
methods. Recalling the image contents of the data sets, these
results are expected. The WY data set covers a smooth rural
area. Although there are elevation differences, no significant
discontinuities and gradient changes exist within a matching
window. Therefore, the mathematical model of the image-
space matching, i.e., approximating the object surface as a
plane, still leads to satisfactory results. On the other hand,
the two other data sets, OSU and SWISS-2, contain many
man-made features which cause discontinuities and gradient
changes in the surface within the matching window. Obtain-
ing a better approximation of the surface and matching in the
object space lead to much better results.
The matching results of all the data sets were also compared
to those of the manual measurements. As was done in the pre-
vious comparison, points that were rejected by'the matching,
were also removed from the adjustment of the manual meas-
urements, in order to allow an objective comparison. The
comparison is shown in Table 6.
For the OSU, SWISS-2 and WY data sets, the results of the
manual measurements are better than the matching results.
Nevertheless, the matching was performed with 60 um resol-
ution images, while the manual measurements were done on
analog photographs (for the OSU and SWISS-2 data-sets)
and 15 um resolution images (for the WY set). For the
SWISS-3, TEXAS and OEEPE data-sets, the results of the
matching procedure are comparable to the manual measure-
403
