w-
n
ig
1-
)
positive effect on the precision. The affine parameters are more impor-
tant for the precision for large windows than for small.
Window Size
Data Set 12 X 12 16 X 16 20 X 20 30 X 30 40 X 40 50 X 59
No AP No AP No AP No AP No AP No AP
Rock 75% 62% | 79% 73% | 78% 75% | 751 00% 717 927 727 827
High1 207. 57 177 97 31 241% 38% 33% 45% 441 51% 50%
Low - - - - 437 351 - - = * + s
The negative effect of using too many parameters is largest when the
window size is small. For the data set “Rock" with good image quality
but large geometric differences between the images, affine parameters
have a positive effect on the reliability for window sizes of 30X30
pixels or higher. In the data set "High1" with lower image quality and
smaller geometric differences affine parameters do have a small nega-
tive influence on the reliability even when the window size is large.
Does Data Snooping have any Effect on the Accuracy?
The parameter set used in the experiments was a 30X30 pixels window,
affine geometric parameters and one multiplicative radiometric para-
meter. All observations (differences between the windows) larger than 3
standard errors were excluded after each iteration, which corresponds
to a 99.7% 2-ended confidence interval of a normal distribution. Exclu-
ded observations could be re-entered in later iterations. The results,
with concern to both precision and reliability, for matchings with and
without data snooping are shown below.
Data Snooping Data Snooping
Data Set No Yes Data Set No Yes
Rock(30) 8.5 um 8.1 um Rock 807 867
Hight 337 367
We could find a small decrease in the root mean square deviation, which
however was not significant on the 5$ level with the Wilcoxon test. The
number of large errors decreased clearly by the data snooping, in spite
of the wide acceptance intervall of the exclusion (3 standard devia-
tions). An optimal threshold could possible give further accuracy im-
provements. This should be further investigated.
Relation Between Window Size and Accuracy
A most important question is the relation between window size and accu-
racy. The practical experiments with different window sizes have been
performed both with and without affine geometric parameters (with a
multiplicative radiometric parameter) as unknowns. The results, expres-
sed as r.m.s. deviations for the used sub sets of data, were adjusted
to the variance component model above with the least squares technique.
The results are shown below. In the figures the root mean square de-
- 581 -
