ng 
he 
on 
to 
SS 
an 
if 
al 
Adaptivity concerning geometrical parameters and window size, as dis- 
cussed by A.W. Gruen (1985), were also investigated. The results were 
not promising. The strategy used gave a lower accuracy, but had the 
advantage of giving a high amount of unrejected points  (Rosenholm, 
1986). 
Computations of the first derivatives from both images, instead of from 
just the mask window, did not significantly increase the accuracy, in 
experiments performed in this investigation. 
SOME COMMENTS REGARDING THE MULTI-POINT MATCHING METHOD 
In this report a multi-point matching method, using finite elements, 
has been presented. The method has only briefly been experimentally 
investigated. However I intend to discuss the properties and the 
possibilities of this method from the preliminary empirical results 
obtained. 
The method is intended for x-parallax (stereo parallax) measurements in 
absolutely oriented, or at least relatively oriented, stereo images. 
Only the x-parallax is used as unknowns. In this investigation a square 
grid was used with a spacing of 8 pixels between the points. Two linear 
radiometric parameters together with 11X11 grid points resulted in al- 
together 123 unknowns in a matching window of 80X80 pixels. The condi- 
tion, minimising the curvature, was used with two weights, 100 and 400. 
With the multi-point matching method we, in this case, use a much smal- 
ler number of pixels per computed point than in single point methods. 
In a 12X12 pixels window each point is computed from 144 pixels, in a 
20X20 pixels window 400 pixels are used, while the corresponding number 
for the multi-point matching method, formulated as above, is 53 pixels 
per point. Each grid point (except for those on the borders) will use 
data from 16X16 pixels for the parallax computation, with the influence 
decreasing linearily with the distance from the grid point in the x- 
and y-directions. However the influence area is larger when constraints 
are used. The method could be expected to have properties diverging 
from any common single point matching method in many ways. What proper- 
ties could the method have? The most relevant comparisons possible to 
make are against the two epipolar matching methods. 
  
  
  
Method 
Data Set No Geom. Persp. MP400 MP100 
Rock(44) 6.7 um 7.0 um 7.4 um 9.6 um 
High1(12) 10.6 um 9.2 um 10.9 um 
  
  
  
  
  
  
  
Root mean square deviations of one-dimensional matchings in the epipo- 
lar direction without geometric parameters, with a linear perspective 
geometric parameter and with the multi-point matching method with the 
weights 400 and 100 using the constraint minimising the curvature 
(MP400 and MP100, respectively). The window size is 20X20 pixels for 
the two single point matching methods 
- 585 -