| purposes 
ience and 
5m colour 
ict that we 
bjects that 
be able to 
ar network 
at this 3D 
nerate the 
| can vary 
lel, giving 
the image 
ige and its 
itputs: an 
Ouse. 
posed of a 
ition DEM 
1). 
reometrical 
ent the 3D 
image. The 
  
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part BS. Istanbul 2004 
  
first produced navigation using a 4 meters DEM clearly shows 
the limits of an approach considering only 2.5D of the space 
(see Figure 2). 
   
  
Figure 2. 4 meter DEM instead of real 3D objects 
The natural solution to this problem consists in considering all 
the objects in a scene as really 3D objects. 
3. NAVIGATION IN REAL 3D 
3.1 Objectives 
The main goal of this work is the visualization of a geometry 
textured with high-resolution images (one meter or better) in 
such a way that the integration of these two complementary 
elements (i.e. images and geometry) improves the perception 
during a virtual overflight of the different objects present in the 
environment. 
  
Figure 3. Real 3D object 
4 
Taking into consideration the 3 dimensions of objects (see 
Figure 3) on top of a traditional 2,5D description of the ground, 
makes it possible to obtain a fully virtual model of a given zone: 
the walls constituting the buildings can be perfectly vertical and 
faithful to reality, the possible concavity of the buildings is also 
considered and visualized (Beck, 2003). 
Furthermore, if the virtual mock-up is precise enough, it could 
be used like a representative model and therefore, it will be 
possible to calculate any desired sight, from any position and 
any virtual model. 
3.2 Source of geometrical data 
Different sources can be used to retrieve useful 3D information: 
e photogrammetric methods (from images, 
laserscanning....) (Ulm, 2003). 
e 3Dsurveying. 
Several papers have presented research in this domain which is 
not the topic of this paper: our point of view is a consumer one 
and we are mainly interested in the characteristics and quality of 
the 3D models retrieved from these different methods in terms 
of capacity to produce a virtual mock-up and allow a good 
quality virtual overflight. 
Note: even if all the processes we describe here can be applied 
to, we do not take into consideration techniques based on 
graphics designer work. Actually, a great part of algorithms 
implemented in this work are used to build missing information 
or to correct data errors. We can suppose that a graphics 
designer will produce an error-free model exactly matching 
visualization requirements. 
3.3 Existing data utilization 
3.3.1 Goal 
Our goal is to generate a 3D overflight of a scene textured with 
any high-resolution image and described by an existing 3D 
model. 
In most cases, this 3D model is totally independent from the 
image (not retrieved from stereoscopy using the image to 
drape): typically, local authorities often own the 3D model of 
their city in order to manage its growth or to manage certain 
risks. These models are often provided bv local surveyors. 
We wish to preserve the possibility of wide navigation areas as 
we have made in navigations on SPOTS data. To ensure the 
continuity of geometry data during the virtual overflight, we 
wish to merge the 3D data with a traditional 2.5D DEM. 
Moreover, we aim at preparing this overflight in a fully 
automatic way. 
3.3.2 Difficulties found 
The heterogeneity of the data that we have previously described 
leads to a non-uniformity problem. Indeed, according to the 
acquirement method employed during the constitution of the 
database, we can find data resulting from: 
e Stereoscopic process 
e Ground measurements operated by surveyors 
* Models drawn by architects 
eii Ete. 
This diversity of methods introduces significant differences into 
the data structures: 
e  Absence/presence of the ground 
e  Continuity/discontinuity of the facetization 
e  Concavity/convexity of the buildings 
e Used facets number for the produced model 
e tc. 
All these differences should not interfere with the process to 
design. 
The Figure 4 shows the detected objects: a building (in grey) 
and two independent structures (in red and green). This example 
shows how a discontinuous acquisition can result in perceiving 
a single building as the union of distinct buildings and thus 
perturb the building identification. Moreover, one of the 
procedures to be carried out deals with the baselines detection 
(see §3.3.3.3). This identification problem will also affect this 
process so that the three different objects will be connected to 
the ground and the final object baseline will erroneously include 
both green and red object baselines.