fields car 
  
trees pylon 
  
people on bridge building 
  
trailer sand 
Figure 1: Examples for different object classes in aerial images 
  
pool in garden car lawn with bushes 
Figure 2: Objects that can be detected easily using color 
with a rectangular mask (diameter 5 pixel), and selection of large 
regions is shown. 
  
raw segmentation 
postprocessed areas 
Figure 3: Transformation and selection of regions after pixel 
classification 
One disadvantage using color is the problem of calibration 
because most images are digitized from pictures. In this case the 
color features of every object class has to be trained for every film 
and every scanner. 
2.2 Multi View 
The extraction of primitives like edges is often incomplete be- 
cause the objects are partly occluded by other objects or due to 
166 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
unfavorable illumination. This problem can be reduced by using 
more than one view of the objects (Roux and McKeown Jr., 1994). 
In the case of aerial images stereo pairs are often available. In 
figure 4, for example, a building is shown in two different views. 
The edges in both images are incomplete. But the combination of 
both segmentations yields a better interpretation with additional 
information of the 3D structure of the building (Haala, 1994). 
      
view 1 view 2 
Figure 4: Edges extracted from two different views 
23 Digital Elevation Model 
One completely different type of input data is a digital elevation 
model (DEM). It can be generated using manual or automatic 
matching of stereo images or by sensors like a laser scanner. A 
DEM is useful for the extraction of objects which are higher than 
their surroundings (e.g., buildings or trees). A popular operator 
for the extraction of high objects is the gray opening. In case of 
noisy data the dual rank, which can be seen as an extension of the 
gray opening, gives better results (Eckstein and Munkelt, 1995). 
The dual rank consists of two successive rank operators. The 
first rank operator is applied with the given rank value while the 
second one uses the “dual” value (i.e., maximum - rank value). 
Therefore the rank value 1 results in a gray opening, and the value 
n (maximum) corresponds to a gray closing. In the case of n/2 
we get two successive median filters. The rank value thus controls 
   
    
  
    
    
   
   
  
  
  
   
   
  
  
   
  
   
    
  
  
  
   
   
   
  
  
  
     
     
      
   
  
  
   
   
     
  
   
   
    
the behavioi 
a rank value 
rank value s 
pixels. 
As an ex: 
a gray imag 
correspondi 
   
gr 
Figure 
At first t 
sponds to tl 
this ground 
6, which is 
objects rer 
extraction 
where the 
The right ii 
detection c 
ing, but the 
Another pr 
to be signi 
  
top h 
Figure 
Shadov 
using the 
problems 
the gray | 
like struc 
is dividec 
pixels car 
other dar} 
sun is sin 
segmenta 
Due to 
be impro 
]. Hlin