CIPA 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey 
specifications: Agfa Rodinal and Kodak Technidol, the second 
one being specific for Technical Pan film. 
Figure 4. Design of the converging network. 
3.1.3 Negative digitisation. The digitisation was done with 
the use of a Vexcel Ultrascan 5000 scanner with a pixel size of 
5 pm. The pixel dimension was calculated in relation to the 
resolution of the APX 100 (150 lines/mm) film and the formula 
proposed by Kraus (1993). The size of the pixel represents 
approximately 1.8mm of the central area of the façade. 
3.2 Photogrammetric process 
The steps followed are the typical ones in the conventional 
photogrammetric process: internal and external orientation and 
DSM Construction. The difference in this case is that the 
Orthobase Pro allows a simultaneous adjustment of these 11 
images. 
The DSM construction is accomplished by automatic methods 
of stereo matching. However, Orthobase Pro cannot take 
advantage of the network as a whole and the models are 
constructed from only a pair of images. Thus, each combination 
of two images allows for a different DSM. The results are 
points that permit the construction of a Triangulated Irregular 
Network (TIN) structure (Peucker et al., 1978). 
3.3 Error control 
The error control was carried out by comparing the z 
coordinates (depth) out of the 75 checkpoints of the DSM 
(Figure 5). The results are expressed as Root Mean Square 
Error (RMSE). The DSM data for a concrete localization (x,y) 
is estimated through interpolation based on the TIN. Thus, it is 
interesting to have a dense TIN, formed by an elevated number 
of points. The contrary is equivalent to a process of 
generalization that would have foreseeable visible negative 
influences on the error statistics. 
The control points will not always be available in their totality, 
since, according to the photographic pair, some may be in 
hidden areas. In this case, the points are eliminated before the 
error control is performed. 
Figure 5. Distribution of the check points. 
4. RESULTS 
4.1 Adjustment error and network design 
The internal orientations were adjusted with errors that were 
always inferior to the pixel (5 pm) size. The external 
orientations were adjusted with standard deviation values of 
0.49 and 0.66 pixels. The calculation of the centres of 
projection had identical results with both films, with variations 
that never surpassed 9 mm with the exception of point 7 (14 
mm). Network adjustments have been carried out from a 
minimum of three images up to a maximum of 11. The 
exactitude of the bundle adjustment was independent of the 
number of exposures included in the network. The standard 
deviation of the adjustment in the minor block (3 images) was 
of 0.48 pixels with 980 tie points. In the 11 image block, the 
standard deviation was of 0.49 pixels with 3400 tie points. 
4.2 Stereo-matching Strategy 
Although Orthobase Pro uses a “simplified structural matching” 
algorithm (Wang, 1998), it is not clear in which stages of the 
process this algorithm or other alternatives take part (Karras et 
al., 1998). The configuration process of stereo matching by the 
user can only be modified in minor factors: the selection of the 
search window sizes and their correlation. It should be noted 
that in epipolar images, the first of these factors might play a 
less important role. However, our tests reveal a surprising result 
that is detailed below. 
The determination of the tie points is done automatically once 
the support points and the size of the search windows and 
correlation are defined. Apparently, the number of tie points 
that are recognized is determined by a threshold value of the 
correlation coefficient. The data associated to the tie points 
include information on two interesting characteristics: the 
number of images in which the point has been acknowledged 
and what those images are. Analyzing the reports, two facts 
may be determined: 1) most tie points are recognized in only 
two images and 2) these pairs are formed by images with the 
minimum convergence angle allowed by the block. When the 
network is formed by 11 images, the majority of the tie points 
are recognized in pairs with a convergence angle of 5°. When a 
network is formed, for example, by images 1-5-9 the tie points