International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
   
   
  
   
  
   
   
   
  
   
   
  
    
    
Figure 9. Example of tiled area 
  
    
[2 
Figure 10. Flight over Toulouse downtown 
4. PERSPECTIVES 
4.1 Performances 
The system described here is designed to enrich the 2,5D 
geometry of a traditional DEM with real 3D data. 
However, we must take into account two significant 
considerations: 
e We wish to increase the geometrical representatitvity 
(and thus the facet number) of the used 3D models, 
e It is essential to guarantee a sufficient frame rate for 
any virtual overflight. 
To achieve these goals, we currently work on a priori 
simplification methods to be applied to 3D objects. 
4.0 Texturation of the 3D model with images 
Currently, we texture the produced geometrical model with an 
orthoimage, which means that we vertically project this image 
on the 3D model: this leads to quasi-uniform walls, which is not 
realistic (see Figure 10). Improving this process means 
resolving two main problems: 
e Data availability 
e Exact extraction of building walls from images. 
Data availability 
Data availability is a requirement which is very difficult to 
comply to: it is essential to have a high number of images taken 
from different points of view to be able to actually see the walls, 
roofs and streets. If we do not have enough images, the only 
possible choice will be between facets with ill- or without 
radiometric information. 
  
Texture extraction 
Supposing that we have enough images to texture all the 
geometrical facets of the model, the next problem deals with the 
radiometric facets extraction process. Apart from the choice to 
be made between the various views of the same facet, it is 
necessary to take into account all used image projections. 
The essential pre-requisite for the wall treatment is the exact 
knowledge of the geometric viewing model for each available 
image. We develop techniques allowing to precisely adjust 
these models with respect to the 3D model geometry: we take 
into consideration methods based on automatic matching. Once 
the viewing model is accurately adjusted, we use it to project 
the 3D model in the image's geometry in order to generate a 
virtual image of the database. This virtual image can be 
perfectly superimposed over the considered image while making 
sure of preserving the faceted structure of the initial model. 
This couple, the projected model and the image, makes it 
possible to obtain all available textures in the given image. This 
process must be applied to all available images for the 
considered scene. 
Once the texture facets are extracted, we will apply the inverse 
projection model to resample all texture facets. This process 
allows to come back to the geometry of the scene and to define 
classic texture coordinates of virtual reality conventions. 
This process makes it possible to generate for each geometrical 
facet as many texture facets as available images. The last step 
will be to choose, for each geometrical facet, the correct texture 
between these various possibilities. 
4.3 Real 3D 
Finally, once the virtual mock up is achieved (meaning modeled 
and textured), we wish to use it not only for virtual overflights 
but also to generate perspective views. Indeed. once the 3D 
scenc is available, it will be possible to define a virtual camera 
system that will be able to produce a virtual image. This image 
will be virtual considering the used camera but real considering 
the used data. Using a virtual system to generate these images 
allows us to simulate any virtual camera. 
We can even consider that the problem of re-sampling any high 
resolution images affected by 3D effects can only be made this 
way in a correct way: this virtual mock-up is an essential step. 
sol 
car 
3D 
nov 
Fin 
for 
WIS 
exti