bof faces before 
. After applying 
orizing the face 
iatching the face 
. The geometric 
the planimetric 
ive based on the 
and the relative 
'e matching was 
ncipal axes ani 
y extracting the 
tner, 1995) ani 
omplicated roof 
ction of the face 
/e edges. 
juare root of the 
nearest neighbor 
r differentiating 
nd extraction is 
| figure (1). The 
  
ond test area 
ssified pixels 0 
The eliminatioi 
. The number d 
classification is 
Michel Morgan 
50 nr. À threshold of 50% is used for the ratio between the interior face pixels and the total number of pixels for each 
segment. ). Two building segments are detected for the first test area and one for the second test area as shown in figure 
(2). Buildings are successfully detected according to the visual interpretation of the DSM and the global comparison to 
the ground truth map. 
Three and two main faces are identified for the buildings in the first test area and four main faces for the building in the 
second test area. Because one of the faces of the first building is outside the ground truth map we excluded it from the 
building extraction processes. The next step is to obtain adjacency information among the roof faces. The outcome for 
the first test area is that each face is adjacent to one face in addition to the background. Each face is adjacent to two or 
three other faces in the second test area. Therefore, rectangle extraction is done for each of the two faces of the 
buildings in the first test area. A height value is assigned to each corner of each rectangle according to the equation of 
the face plane and the planimetric coordinates of the corner. For the building in the second test area, face intersections 
based on the plane equations are done for adjacent faces and line fitting is done for the boundary pixels for the faces 
adjacent to the background. The outline vertices of the building could be calculated by intersecting the outlines. 
Considering the line intersections between faces are more reliable, the outline vertices are obtained by intersecting the 
lines (which are the intersections between the faces) and the outlines. This leads to two solutions for each vertex and by 
averaging their coordinates one solution is obtained for each vertex. The resulting roof faces are overlaid with the 
ground truth map and the detected building segment and displayed in figure (3). 
  
   
  
(a) (b) 
Figure 2. The detected building segments in (a) the first and (b) the second test area. 
x 
22 7 Baskkng faces + 
  
HE i “a red Suit dala XY a par uty dats 
(a) (b) (c) 
Figure 3. The resulting roof faces overlaid with the detected building segments and the ground truth data for 
the buildings in (a and b) the first and (c) the second test area. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 621