The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008 
point cloud consists of approximately 30 percent of these points, 
the result (mean distance) without these points would be 
reduced to less than 6 cm, which is an acceptable result. 
In regions with overlapping datasets, the data had to be fused or 
the overlaps had to be removed to avoid inconsistencies during 
the meshing procedure. Here, the regions were combined by 
removing the less accurate dataset in a particular region. 
Therefore, the manual measurements were used as reference 
data, the contour information was used to fill the occluded parts, 
e.g. the upper parts of the shoulders. 
5.1 Modelling of the surface 
To generate a surface of the full object, the software Geomagic 
Studio was used. First the automatic meshing function was 
activated. This procedure was not successful, because the 
implementation of the software could not handle the line-wise 
structure of the dataset, with a high point density in line 
direction and less density across the line. Some tests were done 
to eliminate this inconvenience under preservation of the high 
frequency information of the dataset. Nevertheless, the result 
contained a number of holes and errors which had to be filled 
and removed. The result is a consistent surface of the object, 
see Figure 8. 
Figure 8. Result of the 3D modelling of the full dataset using 
Geomagic Studio 
6. VISUALIZATION 
The texture mapping was done using one image of the dataset B 
with an in-house developed software. To visualize the extracted 
data with around 600k triangles and to achieve a satisfying 
result the open source software system Blender was used 
(Blender, 2008). It is an open source 3D content creation suite, 
available for all major operating systems under the GNU 
General Public License. Among other features it can be used for 
modelling, texture mapping, animation and rendering, 
comparable with a commercial system, e.g. Maya (Autodesk). 
The model data can be directly imported into the software. 
Using this information the object can be visualized inside the 
software in real-time and virtual reality models can be exported 
in different formats (VRML, COLLADA, etc.). To generate 
movies and high resolution images of the model different 
additional steps are necessary. For example, different 
illumination sources and camera positions, as well as the 
camera path for the camera movement have to be defined in an 
iterative procedure. These steps require, depending on the 
experiences of the operator, some time to become familiar with. 
Because of the lack of texture information in the occluded 
regions, a not fully photorealistically textured model is 
presented in Figure 9. In fact the missing texture was filled up 
by artificial texture patterns. 
Figure 9. Textured model of the Small Buddha 
7. CONCLUSIONS 
We have described the procedures that led to the 3D modelling 
of the Small Buddha of Bamiyan, Afghanistan. We have used 
two old amateur images of the destroyed Buddha figure itself, 
some Sony Cybershot images for the modelling of the now 
empty niche and a contour plot, generated from metric images 
of a past photogrammetric campaign. The local topography 
allowed only images to be taken with very skew angles. 
Therefore many parts of the Buddha figure remained occluded 
in the images and had to be reconstructed from the contour plot. 
Given the difficult data configuration and low quality of the 
primary data, automated methods of model generation do not 
967