- = 
Topologic relationships are recorded through data 
structuring (therefore sometimes referred to as building 
topology). This in fact goes hand in hand with reconstructing 
the objects. To reconstruct a house, we need the ground 
surface defined by a TIN and the outline of the roof (see 
figure 2b). Projecting the roof yields the footprint which 
must be incorporated in the TIN by local retriangulation (see 
figure 2c). This way an integrated representation of terrain 
relief and other topographic objects is assured; the 
topologic relationships between ground surface and solid 
man made objects is established and maintained in the FDS 
database. 
  
Figure 2 Steps of data structuring for 3D FDS. 
Figure 2d indicates that the faces defining the walls are 
oriented to serve 3D visualization which requires that the 
normal vector of each visible face of a solid object should 
point toward the outside of the body. 
  
  
  
  
  
  
Figure 3 (a) Outline of two adjacent roofs, (b) details of the 
roof, (c) two adjacent flat roofs with the same elevation, (d) 
similar to (c) but different elevation, (e) replacing the outline 
of the roofs by their facets after reconstructing the main 
geometry. 
Not all topologic relationships, however, which are shown 
in figure 1 (eg, a node in a body), can be established at this 
863 
stage, because of the lack of information from 
photogrammetric data collection. In some cases adjacency 
can directly be inherited from 2D topology (see figure 3a 
and 3c), but in others (see figure 3d) manual intervention 
becomes necessary. This matter still needs further 
investigation. 
Dealing with occlusions is a bigger problem in 3D than 2D 
topographic mapping. The chances of attaining a tolerable 
algorithmic completion of invisible edges are much less. A 
remedy can be sought in acquiring photography for 3D 
mapping with larger than conventional overlaps, and 
providing better facilities for multi-image access. In a photo 
flight with generous overlap it is more likely that a point 
hidden in one stereomodel is visible in other images. On a 
DP, either local image matching with a 'third party' could be 
attempted, or quickly swapping stereomodels to one where 
the once hidden point can be observed. Despite of 
procedural refinements of aerial photogrammetry, complete 
geometric reconstruction of the urban scene will generally 
not be possible. Since position will usually be defined in a 
national coordinate system, all field completion work can 
unambiguously be linked to the data collected 
photogrammetrically. 
A DXF interface between the FDS database and the 
photogrammetric subsystem for superimposition and 
editing is also convenient for linking up to commercial 
rendering software such as 3D-Studio or virtual reality 
software accepting VRML files such as the 3D viewer WIRL 
on Internet. 
3.2.2 Semantics and Texture 
Digitizing features in a stereomodel is not done without 
interpretation. The semantic information extracted by 
interpretation can easily be encoded by extending the 
codes used for object reconstruction. In model construction 
the codes must be translated to class labels and thematic 
attributes for the FDS database. 
A bigger problem is extracting photo textures, geometrically 
and radiometrically rectifying the image segments and 
storing them in the FDS database. Once the roofs, walls, 
football fields, etc are geometrically determined, the image 
coordinates of the boundary polygons are readily available 
on any photogrammetric plotter. On a DP it is then 
conceptually not difficult to 'cut out' the corresponding 
image segment and geometrically rectify it to the face. 
Before archiving such image segments, it is desirable to 
retouch those images that have alien texture (eg, tree on a 
facade). The problem of occlusions is even more aggravating 
for image mapping than it is for extracting vector data. 
Automatic retouching is not that trivial, but manual image 
manipulation is offered by commercial programmes like 
PHOTOSHOP, PHOTOSTYLER, etc. Another desired 
processing step is radiometric homogenization in order to 
reduce effects of different lighting conditions (see Gruber et 
al, 1995). 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996