The choice of the spacing between the grid points was made without 
previous knowledge of the method, or knowledge of optimal parameters. 
The spacing between the points is probably too small or alternatively, 
which is also indicated in the table, the weight of the ficticious 
observations minimising the curvature and also holding the points 
together is too small. The result should probably have been compared to 
single point methods using smaller window sizes. 
  
  
  
Method 
Data Set No Geom. Persp. MP400 MP100 
Rock 83% 89% 857 857 
High1 36% 36% 50% - 
  
  
  
  
  
  
  
Reliability Measured for the Same Data Sets as in the Previous Table. 
The multi-point matching method gives in the table above a clear 
improvement of the reliability in the "High1" data set but not in the 
"Rock" data set. Why? The answer is most obvious with knowledge of the 
image material. The reasons for gross errors in the data set "Rock" are 
geometrical. Rather many of the check points used for comparision 
between the matchings and the manual measurements used are situated on, 
or close to, pipes. The large window size (80X80 pixels) used by the 
multi-point matching method could not improve the reliability further, 
it could in this case even be a disadvantage as the pipes far away from 
the check point influence the result. 
The data set "High!" on the other hand, has no such extreme geometric 
shapes. The image quality is low and there are large radiometric dif- 
ferences between the images. In such cases the use of large windows is 
a pure advantage, in terms of reliability. Extremely large windows, I 
am thinking of 1000X1000 pixels and larger, could possibly be used, as 
long as the geometric and the radiometric differences are modelled. In 
the multi-point method this is possible. The geometry is well modelled 
by the points with unknown parallaxes. The radiometry could also be 
modelled by bilinear functions, but further research is necessary in 
order to more correctly model the radiometric differences. I would like 
to formulate the following two possible hypotheses about the method. 
The first hypothesis is that the precision of the multi-point matching 
method is related to the grid spacing, and the strength of the 
additional constraints, similar to the relation between the precision 
and the window size in the single point methods. 
The second hypothesis is that the reliability is related to the window 
size. As long as the radiometric differences between the images are 
modelled an increasing window size will correspond to increase in 
reliability. 
This would mean that the method could combine the advantages, and 
eventually the disadvantages, of relatively small windows, concerning 
the precision, and extremely large windows, concerning the reliability. 
A future research task is to investigate the correctness in this hypo- 
thesis and to find optimal parameters. 
- 586 -