ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision", Graz, 2002 
  
  
Figure 5: Reconstructed 3D models of complex buildings 
age of constraints reduces the degree of freedom of some 
parameters, therefore precision of the parameter estimation 
is increased. 
In our building reconstruction system the following types of 
constraints are used: 
e Parameter constraints: establishes a relation between 
two parameters of two building models. For example, 
two building models have the same orientation. 
e Connection constraints. One edge of a building model 
lies on one of edges of the other building model 
e Corner constraints. Two building models share a com- 
mon corner 
e Extension constraints. 
common edge 
Two building models share a 
The constraints are implemented in the least-square adjust- 
ment as weighted observations with standard deviations. The 
weight specifies the strength of the constraint in the adjust- 
ment. 
Evaluating the partition schemes we found that the partitions 
presented in Figure 5 are the best ones. The 3D building 
models were obtained by adding artificial vertical walls to the 
reconstructed roofs. 
A - 360 
The results from the proposed approach are encouraging. The 
method worked well even in difficult conditions where feature 
based approaches would have failed. 
6 CONCLUSIONS AND FUTURE WORK 
A knowledge-based approach for automatic 3D reconstruction 
of buildings from aerial images was presented. The genera- 
tion of building hypotheses in case of large variations in the 
terrain height was described. In this case the building local- 
ization process can no longer be separated from the building 
reconstruction process. Some possible locations of a build- 
ing in the images were determined by combining information 
from the map with image data. These locations were verified 
by the building models generation process. The robustness of 
this method was shown by the presented experiments. The 
correct building model was found for different height values. 
Future work will be directed towards refining the partitioning 
by image information in case all the partitioning schemes are 
rejected by the tree search. 
REFERENCES 
[Bignone et al., 1996] Bignone, F., O.Henricsson, P.Fua, 
and M.Stricker (1996). Automatic extraction of generic 
house roofs from high resolution aerial imagery. In Com-