The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
1009 
(2) 3D modelling using regular surfaces and mesh elements. 
Some geometric details were ignored when we built the basic 
3D model of the prayer-wheel gallery so that the prayer-wheel 
gallery was considered that it is made of planes, smooth 
surfaces and non-smooth surfaces. We discussed a method of 
3D modelling using regular surfaces and mesh elements, its 
processing pipeline (figure 4) is as follow: 
a. Range images are segmented into plane regions, smooth 
surface regions and non-smooth surface regions. 
b. Create a group of cross-section lines from large-scale smooth 
surface regions. 
c. Create another group of cross-section lines, and the two 
groups of cross-section lines are orthogonal. 
d. Generate NURBS surfaces by fitting the cross-section lines. 
e. Create a group of cross-section lines from large-scale plane 
regions, and generate planes by fitting the cross-section lines. 
f. Compute the lines of intersection between neighbording 
planes and surfaces. 
g. Build meshes from the small smooth surface regions, small 
plane regions, and the non-smooth surface regions. 
h. Lastly, get a model closer to CAD applications. 
Figure 4. The pipeline of 3D modelling using regular 
surfaces and mesh elements 
(3) Blocked modeling and merging. In the modelling phase, 
we employed blocked idea to model. The model of each group 
was created, and then the models were merged a whole model. 
Figure 5 is the part of the model of the prayer-wheel gallery. 
3.2 Image data processing pipeline 
In practice, the scanned data is not continuous, although 
contains colour information. 2D images mapping on 3D models 
is satisfactory for 3D digital investigation on diseases of murals. 
The traditional methods are realized by rigidly attaching a 
camera onto the range scanner and thereby fixing the relative 
position and orientation of the two sensors with respect to each 
other [Fr'uh C., 2003, Sequeira V., 2002, Zhao FI.,2003]. Fixing 
the relative position between the 3D range and 2D image 
sensors sacrifices the flexibility of 2D image capture. In fact, 
because of occlusions and self occlusions, the methods above 
described are not suit to the large-scale scenes. We use a hand 
held digital camera to take the images from different angles, in 
different times, in different focal length. It is a technical 
challenge integrating the images from freely moving cameras 
with 3D models or 3D point clouds. Some related works have 
done by [Stamos I., 2008, Zhao W., 2005. ]. I.Stamos’s methods 
assume the existence of at least two vanishing points in the 
scene and register individual 2D images onto a 3D model. W. 
Zhao’s methods align a point cloud computed from the video 
onto the point cloud directly obtained from a 3D sensor. 
Our main goal was to create 3D orthophotoes of murals, 
including (1) generating a big orthophoto for each side wall of 
the prayer-wheel gallery by aligning a sequence of images onto 
the range images and (2) producing 3D orthophotoes for each 
mural by mapping a sequence of images onto a mural model. 
There are some big holes on the range images of the prayer- 
wheel gallery due occlusions. Using the range images are 
difficult to model for full of murals. Thereby, we discussed an 
approach through integrating laser range data with multiview 
geometry for generating the 3D orthophotoes of the murals and 
the orthophotoes of each side wall of the prayer-wheel gallery. 
We use the following method to create the big orthophotoes for 
each side wall. 
(1) Recover multi-view relations from an image sequence by 
structure and motion. 
(2) Compute dense depth map using multi-view stereo. 
(3) Determine the camera poses by aligning 3D point clouds 
from the camera and the 3D sensor using ICP (Iterative Closest 
Point). 
(4) Generating the measurable orthophotoes. 
And then, the pipeline of creating the 3D orthophotoes of the 
murals is shown on figure 1. We need not calculate dense depth 
map, and only need a basic 3D model of the mural from images 
according to multi-view geometry. The basic 3D model aligns 
to the fine 3D model of the mural from high-accuracy laser 
scanning by ICP. 
4. DISEASES MARKING 
Mural survey is an indispensable task before mural repairing, 
and it gives the important data of mural diseases including type, 
distribution, size, etc. Mural survey provides help for the 
scientific and effective restoration of the murals. Early surveys 
of murals, because of material and technical conditions at that 
time, could only use text and simple graphics to describe the 
murals. Recently, the more advanced method of using high- 
resolution digital camera to collect the images is widely applied 
to mural survey. Its processing pipeline is as follow: survey — 
taking photos — mosaic - importing the panoramic image into 
AutoCAD — drawing the base map — layered depicting 
disease. Because mural survey is done in 2D space using this 
method, this method cannot give the true position and size of 
mural diseases. We apply 3D orthophotoes to mark the position 
and size of mural diseases on 3D space.