* The normal matrix is recomputed in each iteration. Thereby it is observed that some of the 
coefficients appear twice in the system 
* The inverse is computed in each iteration for quality analysis of the results 
* Anyor all of the four shaping parameters can be excluded before or during the iterations, if it is 
desired 
* The resampling of the gray values is based on a four term bilinear interpolation, which is 
simplified when the transformation includes only two shifts 
* The convergence criterion depends on an expected height accuracy expressed either in cm 
or in % hg 
< An option for preprocessing (filtering with variable mask size) is included 
* The dimensions of the patches and windows can be rectangular and the windows may also 
have different dimensions 
* The rotation angles can be in rad,degrees or grad and their sequence can be c,0,« or $,0,« 
4. TEST RESULTS 
The results analyzed below refer to the real data. 
4.1 Determination of Weight for Geometric Constraints 
In order to make the constraints fully effective and transform the patches in such a way that the 
expected X,Y ground coordinates are kept, a large weight must be used. If the weight is too 
small, then the parallax is cleared, faster indeed, but the matched patches move unconstrained 
to different X, Y ground positions and the resulting height is wrong. If the weights are too large, 
numerical problems and instabilities can arise in the solution and inversion. Experiments with 
different weights showed that a standard deviation o, 70.001 pixel is appropriate ( with reference 
to the VAX 11/750, 32 - bit real arrays ). A comparison of the behaviour of the transformation with 
standard deviations 6, = 0.001 and 6, = 1000. is shown in Figure 2. 
PARAMETER 7285 / 1.25 1000. 
ALTERATIONS 
1.257 
    
Lis HEIGHT 
XL ..... LEFT X-SHIFT 
YL ..... LEFT Y-SHIFT 
1.07 
0.54 
PARAMETER 7285 / 1.25 / 0.001 
ALTERATIONS 
1 2 13 A 5 6 7 8 9 6 NUMBER OF 
T : ; ITERATIONS 
    
0.0 0.0 
  
-0.5 4 0.57 
-1.0—+ 
  
  
-1.25 + -1.25- 
a) b) 
Figure 2. Alterations of the parameters versus iterations 
a) 04 = 0.001 pi 
b) o,, 21000. pi 
- 289 -