Paolo Gamba 
  
e 24"? + e 242 (3) 
provides a better way to characterize the “differences” between two segments, allowing to discard samples that are 
either too long or too short with respect to the reference. 
  
  
(d) 
Figure 3. An area with both vegetation and buildings: the original LIDAR data is in (a), while the objects 
recognized with a segment similarity threshold of 0.5, 0.7 and 0.8 are respectively in (b), (c) and 
(d). 
The second point to be discussed, i. e. the importance of index sj; for object discrimination, may be more evident by 
looking at some results on a DTM where both trees and buildings are present, as depicted in fig.3 (a). vegetation 
corresponds to an area where the DTM is extremely irregular, the initial data segmentation provides only very short 
segments, and similarity between neighboring segments, in the sense that we defined above, is, when applicable, a very 
small value. In the figure the output of the algorithm changing only the minimum similarity value required to aggregate 
a segment to a plane (either a seed, or a grown plane) is changed, from 0.5 to 0.8. We observe that of course the 
buildings, if decomposed into segments, define segment sets with strong similarity, while the vegetation provides a 
surface where segments may be considered as equally oriented in all directions. When applying larger and larger 
similarity thresholds, only a few of these segments survive the second and third step of the above delineated procedure, 
providing a way to recognize where interesting or useless areas are located. If the threshold is too large, however, also 
  
316 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 
  
- 
e uu u ct mm] 
mn 
th Uf mm oM (8 M 0