Christoph Kaeser 
  
attributes, a respective knowledge/rule database, and processing methods. The processing will proceed from the 
easiest subclasses to the most difficult ones. 
5 SOME RESULTS 
5.1 Buildings 
More details on this part of the project are given in Niederóst (2000). Various cues are used for building detection. 
Niederóst first starts from approximations from VECTOR25, or extracted DSM blobs or a multispectral classification 
(unsupervised K-means classification using 5 information channels). The two last approximations allow detection of 
new buildings, which do not exist in VECTOR25. DSM blobs can be derived with or without use of the DHM25. The 
vector approximation is refined sequentially by a shift, rotation and scaling using scores of all possible solutions within 
a search window, defined by the maximum possible error. Different information channels are involved in the 
computation of these matching scores. At the end each solution's reliability is computed using the traffic light principle. 
First test results from comparison to ground truth showed a detection of 89% of the buildings in the image. 
a 
Figure 3. (a) in red the VECTOR 25 approximation, in green the fit of VECTOR25 to the image data; (b)-(e) various 
cues used in object detection: "artificiality" (factor that separates better man-made objects from natural ones), 
DSM blobs, edge gradient magnitude, edge orientation. 
  
Figure 4. Reconstructed buildings. The colours green (darker) and yellow (lighter) indicate different degree of 
reliability. Green buildings were accepted automatically, yellow after manual inspection (but no editing). 
  
466 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.