Gruen, Armin 
  
Various image analysis algorithms are applied in order to improve planar location, rotation and scale of each single 
building. The solutions are automatically rated to keep the good results and to reject the inaccurate or wrong buildings. 
The building heights are taken from the DSM blobs. For details of the procedure see Niederoest, 2000. 
Figure 5 shows the results of our testdataset Hedingen. For a profound evaluation of the procedure, especially 
concerning the amount of manual interaction, we need more experience with other datasets, which are under 
preparation. 
  
Figure 5. Results of automated house updating (dataset Hedingen) 
For road extraction we have proposed in earlier work the LSB-Snakes for use at small image scales, because of the 
relative simplistic radiometric road model involved. Under ATOMI we are required to use image scales around 
1:15000. This calls for another approach for road modeling. We use for road extraction and updating also multiple cues. 
Finding 3-D edges on the road and especially along the road borders is a crucial component of our approach. These 3-D 
edges are obtained via straight line segment matching in stereopairs. The algorithm exploits the lines’ geometrical and 
photometrical attributes and the geometrical structures. A framework to integrate these information sources using 
probabilistic relaxation has been developed and implemented in order to produce locally consistent matches. The 
proven concept of our AMOBE project for building extraction, to make the transition from 2-D image space to 3-D 
object space as early as possible and to have permanent interaction between the features and cues of these spaces, has 
been observed here as well. Figure 6 gives an overview of the cues used and the algorithms applied. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Feature extraction ; ; 
Stereo color : 3D straight lines 
keriakimages TI Image matching | <> 
recent. 
se s 2 va regions Hypotheses Lp| 3D road 
- Shadows : 
Pl 2p image analysis >|. Road marks, «» ua reconstruction 
Cars, +. verification 
f =) 
VEC25 and —p] Subclass attribute —p] - Road attributes <> 
other input data derivation - Landcover 
- Slope 
  
  
  
  
  
  
  
  
  
Figure 6. Cues used and algorithms applied for road extraction and updating. 
VEC25...Vector data from the given map 1:25 000 
  
314 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000.