International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
  
Un 
Compute radiance images of all texture-images from a 
calibrated response curve. 
6. Re-sample pixels within each triangles based on projective 
transformation between triangle plane and image planc. 
7. Apply filtering and radiometric corrections to the textures. 
8. Create MIP maps; a sequence of textures each of which is 
progressively lower resolution (by a factor of 2) of the 
original image. This is used to solve aliasing problems and 
also for texture size management. 
9. Create bump maps on smooth surfaces, like walls, by 
introducing small variations in surface normals thus adding 
roughness and more realistic look to surfaces when lighted. 
10. Rendering with efficient rendering software. 
  
Figure 3: Textured 3D model. 
Figure 3 shows the final textured model of the frescos. Here a 
brief overview of the different kind of approaches adopted for 
texture mapping and texture correction is reported. More 
details can be found in [El-Hakim et al, 2003]. 
Selecting most appropriate image to texture: With sufficient 
image overlap, the texture of a triangle can be obtained from a 
number of different images. As a rule, the image that has the 
largest texture should be selected, however in order to avoid 
multiple assignments of many different images to adjacent 
triangles, local re-assignment of triangle patches to images was 
employed. In this way the number of triangle edges where 
adjacent triangles are mapped from different images is reduced. 
Texture perspective: The employed method defines a local 
Texel coordinate system for each 3D triangle. Projective 
transformation between the plane of the triangle and image 
plane is determined and used to resample the texture within 
cach triangle. This is followed by low pass filter to remove 
high frequency noise introduced by this re-sampling. The 
geometric accuracy of projective transformation was ensured 
by proper camera calibration, image registration and bundle 
adjustment, which were carried out by TexCapture software. 
Radiometric distortions: Radiometric distortions result along 
common edges of adjacent triangles mapped from different 
images. Though corrections to this problem are still 
experimental, some solutions were adopted for this project. 
Firstly, the response function of the digital camera was 
determined. In this way the non-linear mapping between the 
digitised brightness value for a pixel and the scene radiance 
could be estimated. Such a knowledge allows to merge images 
taken at different exposure setting, different angles, or even 
different imaging devices. Secondly, color corrections between 
adjacent textures obtained from different images are applied to 
minimized the differences. 
Rendering: in order to achieve the desired performance, the 
software for interactive visualisation has been developed at 
VIT [Paquet and Peters, 2002] using scene graph tools. Scene 
Graphs are data structures used to hierarchically organize and 
manage the contents of spatially oriented scene data. They 
offer a high-level alternative to low-level graphics rendering 
APIs such as OpenGL. 
5. LASER SCANNER-BASED 3D MODELING 
Beside photogrammetric surveys, range data of the interior of 
the room were acquired with the Riegl LMS-Z360 laser 
scanning system. Given the relative small size of the room and 
wide field of view of the laser scanner, only 6 scans were 
sufficient to survey the whole volume: 4 for the walls and 2 for 
the ceiling. Setting a high scan resolution value and by placing 
the scanner in the middle of the room a spatial resolution 
(displacement on XY planc) of 5 mm on average was obtained. 
Though the acquiring software, provided in bundle with the 
laser scanner, offers a tool for the scan registration, both this 
procedure and all subsequently ones needed for the 3D 
modeling were performed using Polyworks Modeler of 
[Innovmetric, 2004]. Such a software provides a very powerful 
environment for the interactive modeling of real objects, whose 
geometry has been acquired in terms of very dense point 
clouds. It 1s composed of several modules, which allow the 
user to carry out all the modeling steps, keep the control over 
the entire process and verify the accuracy of the results through 
a number of dedicated tools. In this stage following modules 
has been employed: | 
—  ImAlign for the scan alignment, 
ImMerge for the mesh generation 
—  Iminspect for the model georeferencing 
5.1 Range Data Alignment: 
The interactive manual N-points alignment procedure was 
adopted in this case to register the wall and the ceiling scans 
each other. The acquisition of intensity data greatly helped the 
registration step, as it made easy to recognize matching points 
on the adjacent scans, as shown in figure 4. 
As a result of this first processing step, an approximate 
transformation matrix for each scan pair was obtained and then 
used in the second stage as starting point for the refined 
alignment based on the well-known ICP algorithm. In both 
steps a scan group was locked in order to define the reference 
frame of the model. Then, a global ICP-based alignment 
algorithm was run in order to refine the results of the first 
stage. Such approach [see Soucy et al, 1996] yielded a very 
good registration, with a mean convergence value of 6.3 x 10° 
*. much lower than the preset threshold (10?), and an average 
RMS alignment error of 0.006 m (figure 5). This result should 
confirm the goodness of the registration procedure 
implemented in Polyworks: the residual error is due to the 
inherent accuracy of the laser scanner. Finally, through 
ImMerge, the scans 'were triangulated, in order to model the 
room surface in terms of a unique mesh, which is the most 
suited representation for the model texturing (figure 6). 
   
International 
Jc CITUR 
  
EXT 
EN 
From 
-— 
^ LY! fa La f^ 
Lu 
  
  
  
  
Figure |