/A, 9-11 Nov. 1999 
SEMI-AUTOMATIC 
ONSTRUCTION 
  
3D Primitive 
description 
  
  
Primitive surface 
  
  
  
   
  
  
  
  
         
  
description 
! |... .|. 3D CAD description 
el v (.sat, .dxf) 
ouilding | 
| VR Model 
  
  
  
tomatic and semi-automatic 
the reconstruction algorithm. 
n the right and the flash icon 
tically derived data can be 
' starts by decomposing the 
nitives (rectangles). Each 2D 
esponding 3D primitive. The 
of the 2D primitive applies as 
'emains to be determined are 
y roof type (currently one of 
> building and roof slope. A 
the best fit of the models to 
dels are suitable, the one with 
dplans to rectangles 
  
  
International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 3W14, La Jolla, CA, 9-11 Nov. 1999 
  
  
  
Figure 6: Reconstructed building and DSM 
After this step, the individually reconstructed primitives are 
overlapping 3D solids. They can be output in the form of either 
a list of solid descriptors or a list of planar faces. Most often it is 
desirable to find a building description without overlapping 
parts. As this is a standard CSG problem, a CAD kernel is used 
to perform the necessary merging (Boolean union) operations. 
Finally, a non-overlapping building description is obtained, 
which can be exported an converted into different CAD formats. 
147 
The different steps of the algorithm are exemplary depicted in 
figures 3 to 6. Figure 3 shows the ground plan of the building 
after decomposition in different rectangles. Each rectangle 
triggers the reconstruction of a corresponding primitive ( 
Figure 4). For that purpose roof type, roof height and eaves 
height are estimated by minimizing the differences between the 
DSM surface and the 3D primitive. In order to generate a 
boundary representation, a union of the reconstructed primitives 
is performed in the final step ( 
Figure 5). 
Figure 6 again shows the result of the reconstruction process 
and the DSM surface used for reconstruction overlaid. 
The result of the procedure for a larger area is depicted in 
Figure 7. This test site is located in the city of Stuttgart. It 
covers an area of 1.8 km x 2.3 km and contains 5208 buildings, 
which were reconstructed automatically. For this area the 
groundplans were provided from an digitization of a 1:500 map 
or were derived from the original measurements. This digital 
data set is already available for the complete city of Stuttgart. 
Almost all buildings could be reconstructed by the automatic 
procedure with sufficient accuracy. Gross errors in the 
reconstructed roof shape are only visible for singular buildings. 
These errors were mainly caused by court yards of buildings, 
which were not represented properly in the available 2D GIS. 
Incorrect results can also occur, if parts of the building, like a 
bay or a dormer window are not represented by the ground plan 
and therefore can not be reconstructed by the automatic 
procedure. In this case a manual editing is required in order to 
correct or refine the result of automatic reconstruction. 
For that purpose an interactive tool for the modification of the 
ground plans is available. The tool allows to define, delete and 
modify 2D building primitives. Using the same algorithms as 
those employed in the fully automatic process, three- 
dimensional primitives are reconstructed instantly when the user 
modifies the underlying 2D geometry. In order to support the 
interpretation of the scene by the operator, the editing tool 
enables the simultaneous display of 2D ground plans and 
primitives in an arbitrary number of images, like a scanned map, 
an ortho image or a greyvalue-coded DSM. Beside that, a 3D 
rendered display shows part of the DSM in the vicinity of the 
selected building and the current 3D building reconstruction. A 
more detailed description of the complete algorithm can be 
found in Haala and Brenner (1999). 
4 VISUALIZATION 
The generation of realistic visualizations as the final goal of our 
approach presumes the application of image texture. This is also 
advantageous since the use of natural texture can substitute the 
geometric modeling of building parts at least to a certain extend. 
A realistic impression can for example already be achieved if an 
image of a facade is mapped to the corresponding planar surface 
of the reconstructed building. In this case a geometric 
reconstruction of the different windows is no longer required. 
The mapping of image texture to the corresponding surfaces of 
the reconstructed buildings can for example be realized if aerial 
images are available in addition to the already used DSM and 
groundplans. After reconstruction the wireframes of the 
buildings can be projected into the aerial images and the 
required texture can be extracted and mapped to the 
corresponding surface patches. This procedure, which was also 
applied for the generation of Figure 8 is usually sufficient, if 
image texture is only required for the roofs of the buildings and 
the terrain surface. In order to provide texture for the facades of