ne 
ble 2 
(5 
Table 2 together with Figure 3 show the global reliability and accuracy indica- 
tors. Some interesting conclusions can be drawn from these investigations. Com- 
paring the global indicators RI(T) and AI(T) of the individual network versions 
we notice that good accuracy doesn't correspond necessarily with good reliability. 
Though version B provides for fairly good accuracy, the reliability is poor. This 
refers mainly to the reliability of the x'-coordinate observations. Here we have 
to state that if only two images are available (see also version A) a gross error 
of those image coordinates which belong to the epipolar plane cannot be detected 
at all (zero variance problem - "spur" observations). 
Of course this is not true for control points, so that RI(x") of version A,” which 
represents the average variance of the x'-residuals (with Cus 1 um) is not equal 
to zero. Though the x'-coordinates of version B don't belong exactly to the epi- 
polar plane (647 -63 = 20.4839), their reliability is still poor (that. yields, for 
the non-control points). On the contrary the reliability of the y'-coordinates of 
versions—A,B~ts sufficient (A:R1(y")y = 0.53, B:RI{y') = 0.52), i.e. a gross error 
in those y'-coordinates has a good chance to be detected (see the subsequent 
examples of data - snooping); but here another problem reveals: the problem of 
gross error localization. 
Example C (3 images) provides for better reliability, especially in the x'-coordi- 
nates (RI(x") = 0.43). In this case we are able to' detect a' gross error even in 
the x'-coordinates of non-control points. 
Theexamples D and E show an equally good reliability behaviour in x'- and y'- co- 
ordinates, what is due to the symmetric network arrangement, while the accuracy 
of version E is far more better. 
Summarizing the experiences gained by these investigations and indicated by the 
applied accuracy and reliability indicators we have to prefer definitely the 
arrangement E (images 3-4-7-8), because only in this case the x'- and y"- obser- 
vations are of sufficient reliability together with a satisfactory accuracy of the 
object points. 
Moreover, these investigations may show the necessity of performing a-priori ac- 
curacy and reliability studies for a proposed project. This becomes the more nec- 
essary the more non-transparent the geometrical conditions are. As a rule one 
should aspire reliability indicators which are at least equal or even greater than 
0.6 (RI(x') 2 0.6, RI{y'} 20.6) - inh-accordance with the values gained by the in- 
vestigation of aerial triangulation systems (Grün /11/3. 
Naturally the global indicators don't provide in each practical case for sufficient 
accuracy / reliability of: ald individual object points / observations, e.g. if 
some observations have to be cancelled or if points cannot be observed from certain 
camera stations, whatever the reason may be for that. So the global investigation 
has often to be replaced by an individual checking (see (21)).. To get- further--in- 
sight in detail problems and to become familiar with the data-snooping technique 
the arrangements of Figure 2 and Table 2 will be subject to the data-snooping 
procedure. To keep as close as possible to practice the synthetic image coordinate 
observations of all versions are superimposed by a random generator with means 
= 0 and standard deviations oy:- cy: = 5 um. 
Hy s by" 
Then in all versions A - E different gross errors are introduced as (the gross