‘Complexicity of a priori | 
Performances Knowledge 
    
Manual} 
   
  
  
  
Figure 2 : Systemic parameters 
Which form of interactivity will be introduced? In the 
form of an human operator who choices one corner of a 
building. By this way we hope we could give behavior 
independence between automation and performances and 
thus increase them. 
We develop in details our approach in following parts of 
- this paper. The next onc concerns specificities of our 
approach. Then we explain our detection recognition 
reconstruction methodology. We will present our results 
along methodology explanation in order to be very clear. 
Then we will conclude and develop some perspectives. 
2. OUR APPROACH 
Today, most of problems encountered in recognition 
processes derive from poor performances of detection 
methods. Our approach consists in helping low-level 
process in order to control and understand high-level 
process and to find which modifications we will have to 
do in detection methods to increase performances. As we 
explain in introduction, we decided to tackle this 
particular problem by using interactivity. Our hope is that 
this interactivity could be exchange by an effective 
detection process. In conclusion we will present some 
perspective ideas in this sense. Immediate interest of 
interactivity is to reduce combinational by designing an 
object of interest. Our process follows four steps : 
1- we select manually a building corner, we call it 
seed point in the following, 
2- we detect two first sides of building passing by 
the seed point, 
3- we detect the best parallelogram taking account 
a cost function using the two sides already 
detected, 
4- then we search for homologous parallelogram 
in homologue image. 
We present successively all details of these four steps in 
the next part. 
    
658 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
3- DETECTION AND RECOGNITION PROCESS 
At first, we present area of interest (see figure 3) which 
we used in order to present our methodology. This image 
contains nine buildings which are marked and numbered 
from 1 to 9. We will use these numbers in the following 
in order to compare results. 
  
Figure 3: Area of Interest 
3.1 Manual Seed Point Selection 
Seed point (i.e. building corner) is chosen inside a 
zoomed area of image centered around the object of 
interest in order to localize precisely one of four building 
corners (see figure 4). 
WAAC 
Figure 4: Zoom 
  
Due to manual process selection, during the step 2 we 
will authorize to relax its position into a window of 5 
pixels by 5 pixels in order to compensate bad manual 
selection. 
  
     
  
  
  
  
   
   
  
  
  
  
  
  
  
  
   
  
  
  
  
  
   
   
  
   
     
    
  
  
  
  
  
  
   
  
   
   
  
  
   
  
   
  
   
  
  
  
  
   
   
  
3.2 - 
As $ 
monc 
poter 
on a 
and s 
Sclec 
gradi 
gradi 
equal 
comp 
This : 
Index 
expre 
extrac 
one b 
the sa 
We u 
functi 
show 
two s 
lines | 
classi 
figure 
decliv 
  
Figu