CIPA 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey 
, 
at were 
external 
dues of 
itres of 
iriations 
it 7 (14 
from a 
1. The 
; of the 
;tandard 
;es) was 
Dck, the 
:s. 
itching” 
s of the 
arras et 
\ by the 
n of the 
e noted 
t play a 
ig result 
lly once 
ws and 
; points 
; of the 
: points 
cs: the 
vledged 
/o facts 
in only 
with the 
hen the 
; points 
When a 
e points 
are recognized in pairs 1-5 and 5-9, both of them with 20° of 
convergence. 
These results suggest that the software could identify 
homologous points with large or moderate convergence angles 
but it is designed to use the smallest disposable angles in 
relation to the design of the network. This effect may be caused 
by the less good result obtained when a larger convergence 
angle is used (see 4.3.1). These results may be due to the 
different overlapping of the correlation windows on images 
taken with different angles of incidence. It ought to be 
considered that in this case the homologous points should be 
identified based on the textures of the façade. Other studies 
attempt to resolve these limitations through marking the object 
with a great number of signals (Edmundson & Baker, 2001). 
This strategy, in addition to being very laborious, cannot be 
applied to historic monuments. 
4.3 DSM Precision 
110 DSMs have been constructed (55 DSMs with each type of 
photographic film). These DSMs present different precisions 
according to some variables that have influence on the process. 
Experiments have been performed in order to check the effect 
of these variables and select the best possible combination to 
generate a DSM of greater precision. The variables analysed are 
the following: 
4.3.1 Convergence angle. The precision of the DSM is 
conditioned by the convergence angle of the two images that 
make up the stereoscopic pair. In Figure 6 a collection of 
experiments is represented where it is proven that the error 
increases significantly with that angle: 
Convergence angle 
Figure 6. Relation between the convergence angle and the 
precision of the DSM expressed as RMSE in meters. 
These results indicate the bad behaviour of the application even 
in moderately convergent exposures. This is in apparent 
contradiction with Wang’s work (Erdas web page) where it is 
proven that the turns and distortions are not obstacles for a 
correct matching among the images. It also contradicts the 
theory of the convergent network configuration, whose 
optimum geometry is configured with angles from 60° to 90° 
(Mason, 1994). It is, however, in agreement with the results 
commented in section 4.2 where the recognition of homologous 
points is preferably achieved in neighbouring images with a 
convergence angle of 5°. 
4.3.2 Size of the search windows and correlation. Tests 
have been accomplished to demonstrate that the size of the 
search window influences the number of points of the DSM. 
Different analyses were carried out fixing the coefficient values 
of the threshold correlation (0.85) and the correlation window 
(7x7) but varying the size of the search window. The results 
demonstrate that the tie point number increases and the 
adjustment error descend if the search window is reduced. 
4.3.3 Elimination of gross errors. There are two types of 
gross errors. The first one is due to errors in the identification of 
homologous points and is usual in any photogrammetric 
process. In our case, there is a second type of error derived from 
the existence of hidden areas in some of the images of the pair. 
The first type of error was reduced by examining the 
improvement in the precision of the DSM when the check 
points with the greater errors were eliminated. 
When the improvement curve reaches an inflection point the 
DSM is considered to be purged of large errors. The highest 
value (0.122 m) corresponds to the whole sample. By 
eliminating the greatest error point, the RMSE descends to 
0.105 m. The purging of gross errors finalized when the 5 
points with the greatest errors are eliminated. At this point, the 
curve shows an inflection and the error slope is shallow. The 
values of the confidence interval were estimated according to 
the formula proposed by Li (1991). Figure 7 shows an example 
in which the RMSE value is related to the check points. 
Figure 7. Improvement of the precision of the DSM with the 
elimination of the check points with the greater errors. The 
vertical strokes show the value of the confidence interval at 
the 95%. 
The second kind of errors, derived from false correlations of 
hidden points, were eliminated using geometric criteria since it 
was already calculated which parts of the façade would be 
visible or not for each stereoscope pair. 
4.3.4 DSM errors. Selecting the best values for the search 
and correlation windows and by eliminating gross errors, the 
DSMs are constructed for each pair of stereoscopes with a 5° 
convergence and for each type of film, which may influence in 
the contrast curves and grain size characteristic of each brand. 
The results demonstrate that the average RMSE value with 
Technical Pan (Kodak, 25° ASA) is slightly better than with 
Agfapan APX 100; the DSM present better precision values in 
seven out nine cases analysed. The results are shown in Tablel. 
199