The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
or partially covered with cast shadows (Figure 11). If the grey 
value difference over the building is very high, the algorithm 
cannot extract the building information correctly. 
Figure 11: Highly urban area with big shadow effects. 
4.3 Complex Buildings 
In paragraph 3.2 we proved the capability to extract a rather 
high amount of detail of buildings if and only if there is a clear 
distinction between the regions belonging to the building and 
not belonging to it. 
In cases, were the building is very complex (Figure 12), the 
building outline becomes more and more difficult to be 
identified. In this example even human interpretation is difficult 
and it is hard to decide where the outer boundary of the building 
really is. 
Figure 13 shows an example demonstrating the capability and 
limitations of the discussed approach. The grown region does 
not include the upper terrace and thus it is not included in the 
extracted vectors. The building is incomplete! 
Figure 12: Complex building. 
The problem is caused by the unclear arrangement of “outer” 
walls. Additional constraints and restrictions during the region 
growing process might help to overcome this insufficiency. 
Figure 13: Example of incomplete building extraction. 
4.4 DSM Inaccuracies 
In paragraph 2.3 it has been shown how nDSM information is 
handled in order to determine seed points. One can imagine that 
finding seed points might become difficult if the nDSM is not 
accurate or if it is coarse compared to the building sizes. 
Figure 14 is an extreme example to demonstrate the problem in 
case of a too coarse nDSM. It also makes not much difference if 
the height threshold or kernel size (for homogeneity calculation) 
were chosen differently. A greater height threshold would cause 
the exclusion of buildings for subsequent steps, and a greater 
kernel size would cause regions over small buildings not to be 
considered as building candidates and hence neglected further 
on. Though exaggerated, this example clearly shows the 
possible impact of too coarse DSMs on the quality building 
extraction. 
4.5 Adjacent Buildings 
Adjacent buildings can produce two kinds of problems. The 
first one is shown in Figure 15 where the extracted buildings do 
not share a common border anymore. This misinterpretation is 
due to the fact that building outlines appear as prominent and 
rather thick ribbon-like features in the image. Hence the region 
growing algorithm stops too early and includes only the inner 
edge of these border ribbons. These errors can be removed by 
growing the found regions over their edges until the adjacent 
building has been reached. Of course, a-priori knowledge is 
necessary as one has to know in advance which building blocks 
are compound houses and where borders to adjacent buildings 
are located. 
Figure 14: Wrong seed points due to nDSM inaccuracies. 
723