- 69 - 
b) 
Figure 9. A section of the Crypt, a) synthetic shading replaces 
one of the colour images, b) the proper colour image mapped 
with the technique proposed in this paper. 
The second method , which we ultimately adopted, simply uses 
the manually assigned pairings between subsets of 3D triangles 
and individual 2D images. The only software component 
required here is a simple mapping program that assigns the 
texture coordinates to each vertex of the 3D model by solving 
the collinearity equations used in the photogrammetric module. 
Because the triangles are relatively small, no perspective 
correction was applied to the mapped texture triangles. With 
this method, the original high-resolution 2D images (Figure 8a) 
are always available for processing. Once modified for a given 
task, e.g., like virtual restoration, a simple reload of the VRML 
file in the viewer updates the model. There is no need to re 
compute the projection. The realistic looking nature of the 
model comes from the fact that a calibration of the 2D camera 
guarantees the geometric quality of the mapping; the mapping 
uses all the texture data present in the 2D image, without re 
mapping like in Figure 8b, and a dense 3D model. Therefore a 
user can select the level of resolution for an application and 
then map the texture on a high-resolution 3D model. A section 
of the Crypt is shown in Figure 9a using synthetic shading after 
removal of one colour image and in Figure 9b, the 
corresponding colour image is mapped back onto the surface 
with the technique proposed above. The matching between 2D 
and 3D data is performed interactively. We could include in the 
features detection part of the solution a module that does 
segmentation and matching between 2D and 3D imagery 
(Neugebauer and Klein, 1999). We also do not rely on 
automatic best view computations. If 3D points are occluded 
then at this point we don’t remove the texture mapping around 
that point. Presently we rely on the ability of the user to take 
the 2D snapshots and to pick the best point of views when 
segmenting the 3D model. Another shortcoming is that method 
does not model reflectance. However some of these features 
will be added in the near future. Nevertheless, the results are 
very good especially in a VR room. 
Figure 10. View of the texture mapped model of the Crypt from 
a vantage point that is not physically possible. 
b) 
Figure 11 Example of a virtual restoration, a) current state of 
some of the writings, b) enhanced version with some 
modifications brought to the texture image. Text taken from 
reference (Fonseca, 1979). 
A Byzantine Crypt (rupestrian site - 9th century c.e.), was 
selected as a way to demonstrate the proposed modeling 
method. It is characterized by two entrances one leading to the 
area that served as a cemetery and the other as the church. The 
church portion is divided into two naves according to a 
structure that is typical of the period. The Crypt is characterized