nol 
  
Buildin 
    
  
  
  
  
Building n?6 
  
Building n°7 
  
E Ord 
  
  
  
  
  
  
  
Results are very attractive (sec figure 9). Nevertheless, 
some artifacts appear which can't be resolved either 
because our detection process is ineffective and we have 
to do several efforts to compensate errors, or because is 
due to our monocular approach and consequently this 
kind of error is redhibitory. In building n°2 we extract a 
false side due to bad continuity of gradient sign. In 
building n?9 due to luminances which are equal both on 
building roof and on ground we can't extract its side. 
3.4 - Tridimensional Reconstruction 
As soon as we recognize a building in an image by 
extracting its sides, we look for its homologous in the 
other view using normalized correlation. Maximum of 
correlation gives us disparity value of tested building and 
consequently its elevation. We compare our computed 
results with BDTopo® Data Base of French National 
Geographical Institute, and with manual measures of 
disparities (see table 1). We are under one pixel tolerance 
for major buildings. 
  
n°3 n°4 n°5 n°6 n°7 n°8 n°9 
61 67 64 66 69 66 49 
62 67 65 68 66 66 50 
66,5 66,3 63,8 66,5 66,2 66,3 49,2 
Table 1: Disparities Comparison 
  
  
  
  
  
  
  
  
  
  
  
  
= 
Se 
dii 
  
Figure 9: Closing of Rectangles 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
Line A of table 1 corresponds to our computed 
disparities, line B to manual disparities and C to 
BDTopo® ones. We do not provide disparities on 
buildings n°1 et n°2 because they don’t exist in 
BDTopo®. Some examples are presented here after (see 
figure 10) and then two perspective views of our scene 
(see figure 11 and 12). 
4 - Conclusion 
Our approach presents some interesting aspects. First, it 
is possible to exchange quickly our interactive detection 
by an effective detection process. Nevertheless, this 
interactivity allowed us to realize a complete process 
without integration of low-level errors and consequently 
to better understand difficult points of low-level process. 
Second, the gradient we used (i.e. declivity one) allows 
better detection than classical one and consequently 
improves performances, we are under one pixel of error 
at the end of the process. 
In perspective, we think that a binocular detection will 
better like [Lotti 94] thus it is possible to reconstruct 
buildings which have not horizontal roof. 
    
  
  
Bui 
Bui 
  
  
Bui