REE TOPS 
e models for the 
1e paper. First we 
| comparison of a 
ression of quality 
ork regarding the 
this section. The 
luced in the third 
|. And as a result, 
ment of automatic 
ne. A tree can be 
ter with only one 
he tree top and the 
at means, even in 
etect and measure 
o clicks per object 
ally much of one 
ling or a road. 
tivation that it can 
saving these two 
'S an overview on 
ig" and leads over 
the extraction of 
1 areas. A short 
for the automatic 
In the last section 
are mentioned, an 
is presented at the 
further work is 
utomatic algorithm 
1 99.5% percent of 
ce provider having 
his client with an 
dditional costs on 
s as a marketing 
  
   
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
   
   
c. : 
A 
m 
     
    
C 
= 
“XS 
1 
  
— 
. > i E > ye 2 lá 144 = o x e , . is . . p . . ~ . 
Figure 1: 3D model of the “WelfenschloB” the main building of the University of Hannover "standing" on an aerial image. 
Figure 2: 3D model of the “WelfenschloB” with automatically extracted trees. The map-like image used as ground texture shows the 
result of an edge extractor (A Deriche operator (Deriche 1990) was applied to the aerial image, which is used as ground 
in the figure above. The tree template which was used for the VRML model was originally described in (Saint John 
1997).) 
In this section we want to show, that — even if the applied 
automatic algorithm works not perfect — it could make sense 
to have such an automatic algorithm for the extraction of 
trees. 
The figures on the top of this page should give an impression 
of the impact of 3D tree models in the virtual city model. 
Figure 1 shows a 3D model of the “WelfenschloB”, the main 
building of the University of Hannover. (The reconstruction 
of the buildings in the city was performed manually with the 
inJECT software of INPHO GmbH.) The model was set onto 
an orthoimage, in order to give a more or less realistic 
impression of the scenario. With exactly the same amount of 
manual work one can produce the model which is predicted 
in Figure 2. The map-like texture on the ground was 
computed with a standard edge extractor without any manual 
support. The 3D tree models in the scene are the result of the 
approach for automatic extraction of trees from aerial 
images, which is explained in the following section of this 
paper. 
The situation with the automatic approaches for the 
extraction of topographic objects is, that 99.5% are not 
achieved in the moment. Even a success rate 95% is not 
realistic, except in simple scenes. One of the reasons is, that 
one has to simplify the real world to a model, which in fact 
cannot cover all possibilities. And the more complicated the 
model, the more complex becomes the tuning of the — 
necessarily - increasing number of parameters. As a result, 
the success rate of the most automatic systems is — roughly 
speaking - between 70% and 90%. Significantly lower