Figure 10: Resulting CAD model of the Disentis mon- 
astery 
orientation of the cameras were determined together with 
a subset of the object point coordinates. Due to the two 
different lenses used, two different sets of additional pa- 
rameters were determined accordingly. In the bundle ad- 
justment routine 147 parameters of exterior orientation, 
20 parameters of interior orientation and 771 objectpoint 
coordinates were determined from 2620 image coordinate 
observations. The standard deviation of the unit weight a 
posteriori (69) of the adjustment was 5.5 um, which is ap- 
proximately half the size of a pixel. The theoretical preci- 
sion of the object point coordinates is 1.3 cm (6 y), 1.7 cm 
(6 y) and 1.4 cm (6 7). 
In a second step all object coordinates of the objet model 
are determined by a bundle adjustment routine using the 
previously estimated parameters of interior and exterior 
orientation. This is practically equivalent with a multiple 
ray forward intersection. The 3D-coordinates of 1836 ob- 
ject points were determined from the image coordinate 
observation of 9573 image points in 49 images. In average 
each object point is present in 5.2 images. The accuracy of 
the object can then be estimated by the theoretical preci- 
sion of the object point coordinates, which averages to 1.4 
cm (6 x), 1.8 cm (6 y) and 1.5 cm (67). According to the 
dimensions of the building this is a relative accuracy of 
about 1 : 8'000. 
Beside the geometric accuracy the topologic correctness 
and completeness of the resulting model is of interest. 
Whereas the topologic correctness of the model is already 
supervised by the user during the modelling process, the 
  
276 
completeness depends on the degree of detail and the used 
imagery during the reconstruction process. To demon- 
strate the latter the topologic object description is matched 
with the set of aerial images and with the set of terrestrial 
images separately. The results with respect to the topology 
are given in Figure 11. It is obvious that the aerial images 
allow the reconstruction of most parts of the object, 
whereas the terrestrial images have a significant drawback 
in the reconstruction of the roof parts and some parts on 
the facade, which are occluded. Moreover the geometric 
accuracy of the two subsets suffers from the inferior 
number of image rays and the worse intersection angles. 
Thus the accuracy of the set of aerial images is worse by a 
factor of 2 and the set of terrestrial images by a factor of 3 
than the combined set with all images. 
  
Results from the set of 
aerial images 
Results from the set of 
terrestrial images 
Figure 11: Topologic completeness of the object model 
derived from two subsets of images 
Concerning the degree of detail the results derived in the 
project so far go down to a level, where the windows are 
represented by sets of four straight lines. For other tasks 
or problems a more detailed model might be desirable as 
the final result. Figure 12 shows a few examples how the 
degree of detail can easily be increased. In addition to the 
existing model of the window more details like the sur- 
rounding parts and details in the painting of the facade are 
added. 
Figure 12: Degree of detail on a window feature 
upper row: derived degree of in this project 
lower row: examples for a more detailed 
reconstruction 
This papel 
three-dimée 
tery from ] 
ware env 
processing 
operator a 
standing | 
measurem 
proach ma 
the autom 
tion of the 
ment usin; 
time-cons 
automated 
reliable th 
final solut 
operator i 
pertise, bt 
larger obj 
constructe 
The data : 
work for : 
1998). W 
Markus K 
their thesi 
Benning, \ 
nik zur d 
tendenz« 
Vermesst 
Beyer, H., 
camera-t 
tion No. 
Bürke, B., 
tis. Fests 
Busslinge 
ierung ei 
ma thes: 
Ziirich, 1 
Chapman, 
mote dig 
ronemen 
Dallas, R. 
gramme: 
chine Vi 
Caithnes 
Debevec, 
dering A 
image-b 
Comput 
ley. 199€ 
El-Hakim. 
to flexib 
neering. 
Eos, 1997 
couver, ( 
Gibson, J. 
Mifflin,