ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002 
  
detection and removal of features with low significance to the 
overall appearance of the building. During this feature removal 
step the constrained building model is applied in order to 
preserve the represented building regularities and optimise the 
position of the remaining vertices. 
4.1 Regularization Constraints 
In most cases walls are oriented in parallel to the principal axes 
of the building, which are again often rectangular. It can 
therefore be assumed that most faces of a building model are 
coplanar, parallel and rectangular to other faces in the same 
model. In order to preserve these symmetry relations between 
faces as well as possible during generalisation, these properties 
have to be integrated. As this information is usually not 
explicitly available the constraint building model is constructed. 
Basically it consists of the polygonal building model enriched 
by a set of regularization constraints. The lowest element the 
hierarchy of constraints is the coplanarity constraint, which 
simply groups a set of faces together, each being coplanar to 
any other face in the same set. Two faces are assumed to be 
coplanar if the angle between the normal vectors is close to 0° 
or 180° and their distance is bellow a given threshold. Sets of 
coplanar faces are then again grouped together by a parallelism 
constraint if their faces are parallel to faces of another face set. 
Finally, two or three sets of coplanar or parallel faces are 
grouped by a rectangularity constraint if the faces of each set 
are rectangular to faces in the other two or three sets. 
It is our belief that not every topological relation can be 
detected reliably by an automatic approach. Dependent on the 
quality of the input model, a number of constraints will almost 
always be missed due to errors introduced in the generation of 
the model. Those absent constraints might reduce the quality of 
the final model if missed in high quantities. An application 
should therefore offer the possibility to identify and insert more 
constraints into the constraint building model in a semi- 
automatic fashion to work around those errors and to improve 
the overall quality of the final building model. A semi- 
automatic tool also helps to test the effects of certain constraints 
on the generalisation process by manually adding or removing 
those constraints. Thus, in the current algorithm manual 
improvement of the automatically detected constraints is 
enabled. 
4.2 Model Simplification 
As already discussed, purely geometric considerations are not 
sufficient, if a simplification of objects like buildings is aspired. 
If the geometry of a building model is simplified by arbitrary 
removing vertices or edges, the symmetry of the building will 
irretrievably be disturbed, even if the introduced geometric 
error is small. In order to preserve the regularity of the model 
during generalisation our feature detection and removal 
algorithm allows the use of a manifold set of surface 
simplification operators, each designed to remove one specific 
class of feature types. In contrast to rather simple operators 
used in traditional surface simplification algorithms, our 
operators remove entire features in one continuous process, 
while preserving the integrity of remaining parts of the building 
model. 
Figure 6 depicts the three classes of features which are currently 
distinguished: extrusion, notch and tip. After feature detection 
small features, which are of low importance to the appearance 
of the building are removed by a combination of edge collapse 
and edge foreshortening operations. 
A- 142 
M Rm 
(a) (b) (c) 
Figure 6: Detected features (a) extrusion, (b) notch and (c) tip. 
During feature removal, parts of the building are completely 
eliminated. The optimal shape of the reduced model, however, 
should still be determined by all original points, even though if 
the number of points is reduced in the preceding step. During 
simplification sets of coplanar faces will for example often be 
merged. Nevertheless, vertices which are eliminated from the 
building should still be used to define the overall size of the 
simplified model. In order to calculate the final shape of the 
simplified model the available constraints like coplanarity, 
parallelism or rectangularity between the remaining faces as 
well as all the points of the original model have to be 
integrated. In our approach this is realised by application of 
least squares adjustment. 
4.3 Results 
The algorithm has been implemented and tested on polygonal 
building models of a 3D building dataset. The complexity of 
building models, measured by the number of triangles uses to 
represent the building could in most cases be reduced by more 
than 30%. 
  
Figure 7: Exemplary object model consisting of 2730 triangles 
As an example, the model of the New Palace of Stuttgart 
depicted in Figure 7 originally consists of 2730 triangles. By 
removal of extrusions the number could be cut down to 1837 
triangles. The result of our algorithm to this model is 
demonstrated in more detail in Figure 8 to Figure 11. Figure 8 
and Figure 10 show a part of the original model as it was 
captured from stereo imagery and an existing outline from the 
public Automated Real Estate Map (ALK), respectively. Figure 
11 shows the result of the generalisation process. It is clearly 
visible, that parallelism and rectangularity have been preserved 
for the remaining faces. Especially if the model is textured 
again, as it is depicted in Figure 9, this amount of detail is 
sufficient for realistic visualization even at close distances.