CIPA 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey
system, the point sets from the two programs were directly com
pared, to produce a mean RMS deviation (d). Additionally, they
were compared with a rigid body transformation (giving a mean
standard error Or), to check for small translations and rotations
between the two 3D point sets. Table 3 presents the outcome of
these comparisons according to number of images and ground
control points (GCP).
Table 3. RMS differences (d) and error (Ctr) of rigid body
transformation between point sets after self-calibration
B- PM
B-PMS
images
GCP
d (cm)
CT R (cm)
d (cm)
CT R (cm)
4
6
3.9
2.8
3.0
2.3
4
7
2.8
2.1
3.1
2.0
5
7
2.8
2.0
2.7
2.0
It is seen in ct r that, with the exception of 6 control points, the
point sets indeed remain within the precision of reconstruction,
as presented in Table 2. However, certain significant differences
do exist between the two points sets, seen in the values of d, due
to translation and rotation between the object systems. In fact,
large differences are present even when comparing results from
the same program obtained using different control points, a con
sequence of control point uncertainty. What is to be noted here,
however, is that the two programs are affected differently by the
inaccuracy of the same control points.
A probably more interesting aspect of self-calibration is the re
sults as regards camera geometry, seen in Tables 4 and 5 (in the
second case solutions are given with and without distortion).
Table 4. Calibration results from PhotoModeler
PM
PMS
images
GCP
c(mm)
x 0 (mm)
y 0 (mm)
c(mm)
Xo(mm)
y 0 (mm)
4
6
81.617
0.010
-0.004
81.582
0.012 .
-0.003
4
7
79.211
0.006
-0.001
79.273
0.003
-0.000
5
7
79.374
0.004
-0.000
79.172
0.008
-0.001
Table 5. Calibration results from bundle adjustment (BASTA)
images
GCP
c(mm)
Xo(mm)
y 0 (mm)
k|
k 2
4
6
79.011
78.927
-0.171
0.257
0.118
0.187
6.8x1 O' 7
2.1xl0' 9
4
7
79.011
78.934
-0.219
0.186
0.140
0.204
5.9x1 O' 7
2.0x10' 9
5
7
79.170
79.083
-0.046
0.061
0.308
0.082
2.4x1 O' 6
-5.6xlO' 10
control points with only one or two known object space coordi
nates (the remaining ones are estimated as ‘partial’ tie points in
the adjustments). This particular feature has proved very useful
indeed in architectural applications, where control might not be
available but the regular geometry of the object can be exploited
instead. Considering the example of Fig. 3, one sees that a given
horizontal length L on a planar façade XY allows generating 2
full control points (1,2) and points with known X,Z (like point
6), known Y,Z (points 3,4) or points simply on the plane (point
5). Evidently, additional points on perpendicular planes (known
X or Y or both) may also be considered.
X i ,Y, Z Y, Z Y, Z X 2 ,Y,Z
9: Q O 0
o c
1 3
p 9
4 2
5
c
L
6
o
z x 2 , z
Figure 3. Example for defining ‘partial’ control points by using
one known dimension L and exploiting object geometry.
This approach has been applied in the present case to allow self
calibration, the results of which are seen in Table 6. The known
horizontal length was taken on the left XY façade, giving rise to
2 full (X,Y,Z) control points. One more point was chosen on
the horizontal line (known Y,Z); and three further points on the
façade were also used (known Z).
Table 6. Calibration results from BASTA with ‘indirect’ control
(GCP: 2 full; 1 with known Y, Z; 3 with known Z)
images
c(mm)
Xo(mm)
y 0 (mm)
k|
k 2
4
79.012
78.927
-0.185
0.247
0.124
0.193
6.9x1 O' 7
2.1xl0' 9
5
79.170
79.078
-0.023
0.088
0.303
0.069
2.5x1 O' 6
-6.0xl0' 10
Comparison of Tables 6 and Table 5 reveals that self-calibration
essentially relying on a single known dimension in object space
and making use of certain object properties yields identical ca
mera calibration results to those from full control points. Since
the plots at hand appear, from a purely metric point of view, as
rather questionable, this is probably the approach to be adopted
for the next steps of the present project.
PhotoModeler yields a principal point ‘suspiciously’ coincident
with the image centre (it is not known to these authors whether
the program imposes some internal constraint to principal point
location). The bundle adjustments result in a scatter of the x 0 , y 0
values, generally expected when adjusting non-metric images.
On the other hand, BASTA shows a very strong repeatability re
garding the camera constant; the results from PhotoModeler, on
the contrary, are more scattered, reaching a very large difference
when using 6 control points. The radial distortion estimated by
BASTA differs somewhat in the cases of 4 and 5 images; never
theless, its value is quite small for this normal lens (the calibra
ted curve does not exceed 40 pm at image comers). But if wide-
angle lenses are used, the problem of distortion in PhotoMode
ler must be tackled by partial camera pre-calibration (for simple
approaches see Karras & Mavromati, 2001).
4.2 Use of‘indirect’ control information
Reference has already been made to the fact that BASTA accepts
5. DISCUSSION
Suitable tools exist today for conveniently handling tasks of ar
chitectural photogrammetry, including instances were historic,
usually poorly documented photographs need to be used. Photo
Modeler is such a powerful 3D reconstruction tool. Although
certain of its processes may remain somewhat ‘obscure’, it has
been established here that the reconstruction it provides is equi
valent to rigorous photogrammetric solutions. However, certain
questions must be answered ‘externally’, such as lens distortion
(a considerable problem in several tasks) or the requirement for
more flexible means for tackling unconventional data. It seems
that the combination of commercial program packages with own
software, like the one used here, may prove even more fruitful.
Recently, a ‘new’ set of old images (among them some giving a
bird’s eye view) of the building in question have come to the
authors’ attention. Students of Architecture again had acquired
