moving 
vertical 
index 
| 
| 
| 
v 
formed an 
arying also 
itrol points 
we refined 
] resection 
bed (Figure 
th a double 
atue. Using 
red on the 
ms of the 
eved using 
red on the 
contour plot of Kostka. The image correspondences for 
triangulation were obtained semi-automatically with adaptive 
least squares matching [Gruen, 1985]. The results of the 
bundles are summarized in Table 1, where o, represents the 
standard deviation a posteriori of unit weight and ox,y,z are the 
average standard deviations of the object point coordinates 
located on the Buddha itself and on its immediate vicinity. 
  
  
  
  
  
  
  
  
  
; Oo = 
No. tie ' x Gy o: 
points [m] [m] [m] 
Internet 31 0.016 0.082 0.177 0.330 
Metric 35 0.019 0.076 0.078 0.141 
  
  
  
Table1: Results of the bundle adjustment of the two data sets. 
In the metric block, the control point distribution covered just a 
small part of the images (the statue) while in the Internet block, 
the control information were more distributed on the whole 
images. Figure 6 shows the configuration of the cameras of the 
Internet data set. 
  
Figure 5: A view on the recovered camera poses of the 
Internet images with tie and control points 
2.4 Surface reconstruction 
After the establishment of the adjusted image blocks, the 3-D 
reconstruction of the statue was performed with automatic 
procedures on both data sets and with manual measurements 
only on the metric images. The results of the manual 
measurements will serve for the physical reconstruction of the 
statue. In all cases, Wallis filtering was applied to remove the 
low frequencies and enhance the texture information in the 
images. 
2.4.1 Automatic reconstruction from the Internet images 
A multi-image geometrically constrained least squares 
matching software package developed at our Institute was 
applied on the Internet images [Grün et al, 2001]. The 
automatic surface reconstruction works in fully automated 
mode according to the following procedure: 
1. Selection of one image as the master image. 
2. Extraction of a very dense pattern of feature points in the 
master image using the Moravec operator. At this stage, the 
master image is subdivided into 7 x 7 pixel image patches 
and within each patch only one feature point which gains the 
highest interest value is selected. 
3.For each feature point, using the epipolar geometry 
determined in photo-triangulation, we get the approximate 
matches for the following MPGC (Multi-Photo 
Geometrically Constrained) matching procedure by standard 
cross-correlation technique. 
4. MPGC is applied for fine matching, including patch 
reshaping. MPGC exploits a priori known geometric 
information to constrain the solution and simultaneous use of 
more than two images (Gruen, 1985; Gruen et al., 1988; 
Baltsavias, 1991). 
The difficulties of this data set lie in the large differences 
between the images, due to the different acquisition time, the 
illumination conditions and the different image scales. 
2.4.2 Automatic reconstruction from the metric images 
The 3-D model of the Buddha statue was generated with 
VirtuoZo digital photogrammetric systems. The matching 
method used by VirtuoZo is a global image matching technique 
based on relaxation algorithm (VirtuoZo NT, 1999). It uses 
both grid point matching and feature point matching. The 
important aspect of this matching algorithm is its smoothness 
constraint satisfaction procedure. With the  smoothness 
constraint, poor texture areas can be bridged, assuming that the 
model surface varies smoothly over the image area. Through 
the VirtuoZo pre-processing module, the user can manually or 
semi-automatically measure some features like ridges, edges 
and regions in difficult or hidden areas. These features are used 
as breaklines and planar surfaces can be interpolated e.g. 
between two edges. In VirtuoZo, first the feature point based 
matching method is used to compute a relative orientation 
between couples of images. Then the measured features are 
used to weight the smoothness constraints while the found 
approximations are used in the following global matching 
method (Zhang et al., 1992). In our application, images B and C 
of the metric data set were used to reconstruct the 3-D model. A 
regular image grid with 9 pixels spacing was matched using a 
patch size of 9 x 9 pixels and 4 pyramid levels. As result, 356 x 
311 3-D points were generated. Due to the smoothness 
constraint and grid-point based matching, the very small 
features of the dress were filtered or skipped. Therefore these 
important small features had to be measured manually. 
2.4.3 Manual reconstruction from the metric images 
The dress of the Buddha is rich of folds, which are between 5 
and 10 cm in width (Figure 6). 
  
Figure 6: The folds on the dress of the Great Buddha that are 
reconstructed with manual measurements 
Only precise manual measurements can reconstruct the exact 
shape and curvature of the dress. Therefore the metric images 
are imported to the VirtuoZo stereo digitize module (Virtuozo 
NT, 1999) and manual stereoscopic measurements are 
performed. The three stereo-models A/C, A/B and B/C (Figure 
3) are set up and points are measured along horizontal profiles 
of 20 cm increment while the folds and the main edges are 
measured as breaklines. 
—365—