CIPA 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey 
464 
camera (with each lens that should be used) once before 
the campaign, unless it is subject to unexpected physical 
impact. 
Figure 1: Calibration target for camera calibration. The red 
pyramids are the camera positions from an exemplary cal 
ibration. In the center of the target one can see the marker 
for automatic registration of the circles. 
2.2 Data Capture 
The actual recording consists of taking hand-held pictures 
of the object, as it is nowadays done for any close-range 
application. In order to provide sufficient redundancy for 
reliable automatic reconstruction it is however necessary to 
make sure that all relevant parts of the object are visible in 
at least 3, better 4 images (rather than the theoretical min 
imum of 2 for reconstruction from calibrated views). Fur 
thermore we recommend an overlap of 80% (rather than 
60%), which helps to ensure good image matching results 
at no extra cost (except for longer computation times). Fig 
ure 2 shows an example of a recording sequence captured 
in Sagalassos. 
3 RECONSTRUCTION 
(b) 
Figure 2: Recording for automatic reconstruction, (a) 
Four images of a piece of the ’dancing girls frieze’ from 
the Augustan ’Northwest Heroon’ in Sagalassos. (b) 3- 
dimensional view of the recording setup. 
3.1 Preprocessing 
Before starting the actual reconstruction process lens dis 
tortion is corrected, i.e. the images are warped to the ideal 
pinhole camera geometry with the distortion parameters 
determined through camera calibration. Furthermore the 
user interactively has to segment the object to be modeled 
from the rest of the image by digitizing its outline. The 
segmentation is the only user interaction required for 3D 
reconstruction apart from defining the sequence (see next 
section) - it cannot be automated, because it is a decision 
of the user, which cannot be derived from the scene ge 
ometry alone. Note however that digitizing only needs a 
few mouse-clicks, if it is assisted by the life-wire method, 
as it is implemented in any image-processing toolkit (see 
Figure 3. 
3.2 Matching and Orientation 
Before homologous points are extracted for image orien 
tation, the images are partitioned into regular tiles to en 
sure a good coverage of the object for orientation. We 
use a partitioning of 8 x 6 tiles (which leads to roughly 
squared tiles for standard cameras). A number of homol 
ogous points is extracted for each tile with a hierarchical 
area-based matcher (Klaus et al., 2002). The same points 
are used for a set of 5 consecutive frames of the sequence to 
achieve a stable image block. In our experiments we have 
manually ordered the images (see Figure 4). In principle 
it is possible to match all possible pairs of images and de 
termine the ordering automatically using the homologous 
points, if one is willing to accept the considerably higher 
computational cost. 
With the homologous points an initial estimate for the rel 
ative orientations is sequentially recovered. Since the au 
tomatically matched correspondences can contain a large 
percentage of false matches (in some experiments up to 
30%), it is necessary to use a robust estimation method. 
The standard algorithm for the task is RANSAC (Torr and 
Murray, 1997). Then the approximate 3D points are com 
puted by forward intersection and the final orientations are 
estimated via bundle adjustment with iterative outlier elim-