ISPRS Commission III, Vol.34, Part 3A , Photogrammetric Computer Vision“, Graz, 2002 
  
For a gable roof primitive, the height of the ridge is considered 
as the height of the reconstructed 3D top line if the top 
lines were detected in both images and the 3D line could be 
reconstructed. Otherwise, in case of of a symmetric gable 
roof, the approximate positions of the projected ridge in the 
images can be deducted by taking into account the symmetry 
of the gable roof. Then the 3D ridge can be reconstructed by 
matching these two approximate line segments. In case of a 
non-symmetric gable roof if no image lines are found, default 
values are adopted for the roof parameters. 
To improve the parameters of a building model, the hypoth- 
esized model is fit to image data. The edges of the projected 
wire frame of the model are fit to gradients of the pixels 
from both images simultaneously. The fitting algorithm is 
described as an iterative least-squares algorithm. It estimates 
the changes of the parameter values that have to be applied 
in order to minimize the square sum of the perpendicular dis- 
tances of the image pixels to the nearest wire frame edge 
([Suveg and Vosselman, 2002]). 
The resulting 3D building models are evaluated by computing 
a score function based on the formulation of the mutual in- 
formation between the building model and the images. This 
score combines two measures. One of them counts for evi- 
dence along the projected building model contour in the im- 
ages. If the hypothesis is correct, we expect to find changes in 
image gradients along the projected model contours. The sec- 
ond one measures image intensity similarity over two images. 
Given an object hypothesis and a pose, a point to point map- 
ping can be defined between images. If the hypothesis is cor- 
rect then the intensities at corresponding pixels will be highly 
correlated. The building model with the highest score is taken 
as the correct hypothesis ([Suveg and Vosselman, 2001]). 
5 RESULTS 
The test data consists of high-resolution aerial images. The 
scale of the images is 1:1300 and they are scanned at 15 
mikrons. Two images with 60% overlap are used. The inte- 
rior orientation parameters of the camera and also the exte- 
rior orientation parameters of the images are known. A 2D 
GIS database containing the ground planes of the buildings 
is given. 
In our current implementation five hypotheses are generated 
corresponding to a flat roof building primitive and two sym- 
metric and two non-symmetric gable roof primitives with dif- 
ferent orientations. Therefore we can reconstruct only flat 
roof buildings, gable roof buildings or buildings formed by 
combining these two building types. However the building 
library can be easily extended with other primitive building 
models. 
The first experiment was to generate and evaluate building 
hypotheses for simple buildings composed by only one build- 
ing primitive. First the possible locations of the building in 
the images are determined. Afterwards, each of these pos- 
sible locations are verified by generating building hypothe- 
ses. The building hypotheses derived from outlines of building 
footprints from the map are generated corresponding to the 
building models from the building library. Next, the building 
hypotheses are fit to the image data. The scores computed 
for matching the hypotheses against the images are used to 
choose the best model. 
The resultant building models projected back into one of the 
  
height | flat | gablel 
gable2 | nonsymg2 
223 -249.1 | -376.6 | -300.8 -193.4 
231 -236.7 | -279.5 | -308.8 -154.0 
235 -288.1 | -299.7 | -99.6 -173.8 
250 -186.8 - 138.7 - 
254 -214.8 - 37.4 - 
  
  
height | flat | gablel | gable2 | nonsymg2 
220 -242.6 | -324.3 | -369.0 -254.1 
224 -230.2 | -226.9 | -228.1 -154.3 
250 -168.2 | -230.6 | -265.2 58.7 
255 -162.9 | -267.9 | -295.4 -65.3 
Figure 4: Reconstructed 3D models of simple buildings 
images are presented in figure 4. The tables show the scores 
computed for different height values and for 4 building prim- 
itives (flat roof, two symmetrical gable roofs with different 
orientations and a non-symmetrical gable roof with the ori- 
entation along the larger edge of the building base). If a gable 
roof model oriented along the larger edge of the building is 
found then the search stops and the other gable roof is not 
generated at all. This is the case of the first building from 4, 
when for the last two height values only a gable roof model 
is generated. 
Next, we tested our approach on complex buildings. First, the 
partitioning of the building into building primitives based on 
the ground plan is performed. Then, for each resultant build- 
ing part, hypotheses are generated and the best fit hypothesis 
is selected as building model. The final CSG tree describ- 
ing the building is further refined by simultaneous fitting of 
the building models contained in the CSG tree. Constraints, 
which describe geometric relationships between building mod- 
els, are incorporated in the global fitting algorithm. The us- 
À - 359