6. Classification of gross errors: 
Regarding the different effects of gross errors related to their size we can group them into 3 
different classes: 
1. small gross errors 
2. medium-sized gross errors 
3. large gross errors 
The classification bounds are not fixed, they depend onto the geometry and may vary for different 
photogrammetric blocks. 
All gross errors greater 4:g and less 50'c can be designated as small gross errors. They have no 
significant influence onto the orientation of the models and do not disturb the domain of lineari- 
ty of the adjustment. Gross errors of the stochastical model and systematic errors are not taken 
into account but can be considered as small gross errors. Errors less than 4:0 are integrated 
within the random errors. 
All errors between 50-c and 2-3 base lengths belong to the medium-sized gross errors. They have 
no big influence onto the geometry of the photogrammetric block and don't disturb the convergence 
of the adjustment but they are not within the range of the linearization and the solution may tend 
to a different O-point. Errors bigger than 3 base lengths are named large gross errors. They change 
the geometry of the block severely and cause worse convergence or even divergence. Especially for 
blocks with bad geometry the adjustment must be stopped before reaching the point of convergence: 
7. Location of small gross errors: 
T 
The location of small gross errors poses no problems for the robust adjustment with the chosen 
weight function. Even small gross errors at the limit of location are detected as long as the ob- 
servations are sufficiently well distributed within the models. 
Only for really worse distributions the consideration of the local redundancy (see formula 2) 
would improve the effectiveness of the procedure. The check for the inherent limit of localization 
can be performed only with artificiat data. Example 1 shows that the introduced errors greater than 
the lower limit of 50 are located without any wrong decision. This lower limit is even better than 
the theoretical expectation for the statistical test. (R. Schroth, 1980). 
Example 2 shows a practical photogrammetric block and is demonstrating the effectiveness of tne 
robust estimators. At first data cleaning has been performed manually and the cleaned data have 
been submitted to the automatic procedure. Although the residuals after the manual procedure did 
not indicate remaining gross errors, the automatic procedure located further ones. 
8. Modifications of the procedure with respect to medium-sized and large gross errors: 
Medium-sized gross errors and all the more large gross errors do not belong to any normal distri- 
bution of observations, they are independent from the a priori weights introduced into the ad- 
justment. Thus as long as bigger gross errors have still an influence onto the adjustment al] 
photogrammetric observations are treated with the starting weight 1, used as a priori weight in 
the weight function. This starting weight tends to the introduced specific a priori weight in de- 
pendency on the value of Q by a weight function. The same is true for all non-photogrammetric ob- 
servations, but for them the starting weight 1/100 is used. The weight function is as follows: