SESSION 10: 
  
DIGITAL SURFACE MODEL GENERATION USING IMAGE MATCH- 
ING TECHNIQUES 
  
  
CHAIRMAN: O. Hofmann (FRG) 
AUTHOR: D. Rosenholm (Sweden) 
TITLE: ACCURACY IMPROVEMENT OF DIGITAL MATCHING FOR EVALU- 
ATION OF DIGITAL ELEVATION MODELS 
DISCUSSION: 
Grün (Switzerland): I think it is a very important extension of the 
Fórstner (FRG): 
  
Rosenholm: 
Rauhala (USA): 
  
normal matching techniques to switch over to, like 
we call it, multi-patch matching. If you add geo- 
metrical constraints, you get a combination of what 
Baltsavias has reported yesterday and multi-patch 
matching, then you end up with a very powerful 
System. Just one note. We have proposed a similar 
method for multi-point matching. It does not only 
consider the parallaxes to be constrained but it 
works similarly to a planimetric block adjustment. 
It considers the patches as models, works with tie 
points, and actually executes constraints between 
the transformed patches, between the deformed 
patches if you use the shaping parameters. So 
it not only constrains parallaxes but all the other 
shaping parameters. Clearly, I think multi-point 
matching is a very important aspect and it has very 
good prospectives. 
From your investigations on the precision of the 
matching one could conclude that you could use two 
or even three times larger pixel size still getting 
the same accuracy in microns. Your curves of the 
dependency of the window size and the standard 
deviation are in full agreement with theory, and 
usually you should get an accuracy of about a fifth 
Of a pixel. And if you get a lesser accuracy you 
can increase your pixel size, and that means you can 
use less pixels which speeds up the whole computa- 
tion. What pixel did you actually use? 
25 microns. And I agree with your comment. 
I am very much interested in your study because it 
is along the lines of global least squares corre- 
lation where I use array algebra for the numerical 
solution of the finite element model. And I guess 
computing will be your main practical problem. 
The method is very computing intensive especially 
to get the speed of highly automated systems. You 
88