Liang-Chien Chen 
  
  
Color Image (Left) 
| 
| Color Image (Right) r1 
J 
  
dt 
| Epipolar Transformation 
  
  
  
E 1*1 
  
  
  
  
[ | 
| Epipolar Image (Left) | Epipolar Image (Right) 
[ | 
Segmentation by Computation of Edge 
Region Growing Strength 
  
  
  
  
  
Y 
Preliminary Selection for 
Building Blocks 
  
  
  
  
  
* 
Detection for Precision Buildings and Corners | 
| 
Y 
| Image Matching | 
Y 
| Building Boundaries with Disparities | 
  
  
  
  
Figure 1. Flowchart of the Proposed Scheme 
2.1 Epipolar Transformation 
In order to exclude the vertical parallax to simplify the successive processing, we transform the image coordinates 
to an epipolar system. After the determination for orientation parameters we redefine the axis in coincidence with 
the flight direction. Then the epipolar images can be generated by an orthorectification procedure provided that a 
ground surface is available (Chen & Lee, 1993). Thus, the buildings will yield significant horizontal parallax. 
2.2 Segmentation 
In order to segment buildings from backgrounds in an image, one assumption of being made is that the roof of a 
building is radiometrically homogeneous. Two approaches are possible. The first is the classification (clustering) 
procedure. The second one is called region growing. The problems of using clustering approach encounter 
difficulties of selecting number of clusters and lack considerations of local features. Thus, region-growing 
approach is used in this investigation. 
Referring to fig.2, the method of region growing used in this investigation is stated as follows (Dang, et. al., 1994). 
We first select a seed point then the following procedure will apply to the whole image. The grey value of point S 
is compared to its neighborhoods (P,-P;) band by band. If the grey value difference between S and P , is less than a 
threshold for each band, then the two points are merged into a region. The region begins to grow until the stability. 
The growing procedure is then reapplied to the points which have not been treated. The method produces unique 
result when different sequences of seed points are selected. 
  
  
  
P, Pa] Ps 
P, S P, 
P; P, P, 
  
  
  
  
  
Fig.2. Illustration of a Seed Point and Neighborhoods 
  
170 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.