, but 
| field 
in of 
ial data. 
"ied out 
aka l'aval 
ne 
ially 
areas 
ufficient 
points. 
oto- 
triple 
er to 
y be 
ethods 
into 
je 
statistical 
nall gross 
nly one 
and very 
le gross 
or the 
ter programs 
practical 
erformance 
simulation 
e data 
do the 
iny method 
ird to the 
|justment 
| particulary 
ability of 
ting small 
ly chosen 
   
13 
as it reflects the prevailing practical situation. Admittedly the chosen examples were somewhat 
extreme which was critisized by some participants. 
The data sets of phase II of the experiment were established in such a way that the performance 
of error detection procedures could be evaluated in particular with regard to the separation of 
multiple small gross errors from random and small systematic errors. 
The experiments and their results are described and commented in detail by W. Forstner in [5| and 
|6| and need not be repeated here. However, the following general statements can be made about 
the results: 
- the methods of error detection applied cover a wide range, also the effectiveness of methods, 
the success rates, and the degree of experience vary widely 
- pre-error detection procedures or robust methods are necessary for the identification of large 
gross errors 
- automatic elimination algorithms are generally more economic than manual or interactive pro- 
cedures; the total number of runs is considerably less 
- in the final stage error detection methods which are based on or are more or less equivalent 
to a statistical test (data snooping or something similar) give generally best restults. 
It can also be noted that in the first phàse some participants overestimated their results con- 
siderably. This attitude changed noticeably in phase II. Thus obviously a learning process took 
place, and the evaluation of error detection algorithms has become quite realistic. 
Evidently, the experiments of the WG have simulated the further development of gross error de- 
tection algorithms and have given deeper insight into the problems. It is now generally recog- 
nized that algorithmic solutions are not only possible but that they are most successful in 
ordinary*cases, i.e. when the reliability rules for project planning are observed in aerial tri- 
angulation. Thus, a highly important stage of development has been reached, and it is expected 
that such error detection programs will be generally applied in practice. 
It has also been confirmed, however, that in extreme cases, when multiple and large gross errors 
go together with poor geometric stability of blocks, any error detection algorithm, whether auto- 
matic or manual, may break down or not find a proper solution. 
4. Refinement of the stochastic model 
4.1. The successful development in aerial triangulation with regard to the elimination of 
systematic and gross errors and the related high precision level obtained has opened the view 
for the general problem of the mathematical model of photogrammetric point determination. 
Block adjustment with additional parameters clearly constitutes a refinement of the functional 
model. Gross observational errors can be considered as relating to the functional model or to 
the stochastic model, depending on the approach. However, the stochastic properties of aerial 
photographs and image coordinates have not been investigated thoroughly, So far. And no 
attempt has been made to take correlation between image coordinates properly into account in 
‚block adustment programs. 
Previous investigations,|8| for instance, have clearly established that image coordinates within 
photographs and between photographs are considerably correlated. The magnitude of correlation 
naturally depends on the extent to which systematic image errors have been taken out.