The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
1080 
The Surphaser* data should be acquired from a tripod at a 
height not exceeding 2.5m due to the windy conditions. This 
results in the scanning not reaching the top sections. Therefore, 
a long-range TOF pulsed scanner, Leica HDS3000*, was used 
to fill-in those gaps on the top of the structure by placing the 
scanner on higher grounds at 80-100 meters away. Figure 7 
shows top of the Caryatids porch which was not visible from 
the close-range Surphaser* but visible to Leica* scanner. 
For texture mapping, HDR images were taken. This requires 
taking at least 4 images at different shutter speed and 
combining them to create one HDR image. 
3.2 Range Data Processing 
Processing of the data was performed with commercial as well 
as our own in-house software tools, which were developed to 
achieve high geometric accuracy and visual quality while 
increasing the level of automation. However, an amount of user 
interaction and editing is still unavoidable. The raw scans, 
which are collections of XYZ points in the scanner coordinate 
system, contain errors and noise that must be filtered out and 
holes that should be filled (Weyrich et al., 2004). Next step is 
the aligning or registration of all the scans in one coordinate 
system. Due to object size and shape and obstacles, it is 
necessary to use a large number of scans from different 
locations and directions to cover every surface at the desired 
spatial resolution or level of detail. Aligning those scans 
requires significant effort and affects the final accuracy of the 
3D model. It is performed in two steps: (1) initial alignment 
using positioning device or the data itself by selecting common 
points between the scans; followed by (2) a more precise 
Iterative Closest Point (ICP) technique (Salvi et al., 2007). A 
global alignment is done at the end to minimize and distribute 
remaining errors equally. We perform the first step in the field 
on a 64-bit notebook PC with 4 Giga Bytes of RAM. As soon as 
a scan is completed, it is first simplified to 2% of its original 
size for faster processing then three common points are selected 
and used for initial alignment with the preceding scan. This is 
done while the scanner is acquiring a new scan, so it does not 
consume additional time. We also use this to ensure full 
coverage before moving the scanner to the next position. Once 
the scans are aligned, they need to be integrated to remove 
redundant points in the overlap region followed by the 
reconstruction of a triangular mesh that closely approximates 
the surface of the object (Varady et al., 1997). 
After the mesh has been reconstructed, some repairing is often 
needed to fill cracks and holes and fix incorrect triangles and 
degenerate or non-manifold surface parts (Borodin et al., 2002, 
Liepa 2003). Such errors result in visible faults, and lighting 
blemishes due to incorrect surface normals. Another problem is 
the fact that the object is rarely sampled optimally. Some areas 
such as edges and high curvature surfaces are usually under 
sampled and end up joined by a transitional surface rather than 
a sharp edge, while flat areas are often over-sampled. For 
accurate documentation and visual realism, edges and sharp 
comers must be accurately preserved in the model. Dey et al., 
2001 proposed a technique for automatic detection and 
correction of such sampling problems while Luebke et al., 2002 
survey simplification techniques needed to deal with over- 
sampling. Surface subdivision is another way to improve under 
sampled areas (Zorin et al., 1996). The triangles in these areas 
are subdivided into smaller triangles with points shifted 
according to pre-set rules. Other methods to sharpen edges in 
meshes are available (e.g. Lai et al., 2007). 
3.3 Texture Mapping 
Colouring and texturing was captured with the Canon* 5D 
digital camera, a 12M Pixels full-frame SLR camera. We create 
HDR images from the captured multi-exposure raw images, as 
mentioned in section 3.1 above. The images are registered with 
the geometric model using common points between them and 
the 3D model. In effect, this is finding the camera pose using 
the model points as control points. This must be done for every 
image unless the camera is fixed to the scanner, then it may 
only be done once. However, mounting and fixing the camera 
to the scanner means that the images are taken at the same time 
and location as scanning. This is not necessary the best for 
texture images since we need to select the time of day that 
provides the best lighting, take images in a short period of time 
to ensure small lighting changes, and select the best distance, 
viewing angle, and camera setting. Thus, we opted for the 
taking the texture images independent of the scanner, which 
necessitated the development of an automated approach that 
registers and calibrates each image with the 3D geometry. One 
technique to facilitate this operation registers the texture images 
together first then use only one of the images to register with 
the geometry (El-Hakim et al, 2004 and Stamos et al, 2008). 
For this approach to be accurate, it requires taking images with 
sufficient overlap and strong configuration, which imposes 
restrictions in the field. We use this approach only as an initial 
estimation for a more accurate registration based on matching 
of features in the texture image and the scanner intensity image. 
Once the images are registered with the geometric model, 
several geometric and radiometric processes have to be carried 
out to ensure seamless transitions and distortion-free texture 
maps (El-Hakim and Beraldin, 2007). 
Figure 7. Two snapshots of the compressed 3D model