d
-—
ted to the actual methods and image materials. In images of low radio-
metric quality and with small geometrical differences (High1) the
optimum window size is large compared to images where the radiometric
quality is good and the geometrical differences are large (Rock). The
optimum window size is larger when using additional geometric para-
meters than when not. It is also possible to see that moderate deviat-
ions from the optimum window size are of small importance for the
precision.
One interesting aspect is the size of the constant method dependent
variance component, which is not influenced by the window size. This
variance component is the most difficult to explain. It is not probably
that it could be solely explained by errors in the manual measurements,
which would thus correspond to be between 7 and 9 um. The methods for
manual measurements were designed to reduce different errors. The pos-
sible explanations have to be found in differences between the methods.
This questions is not answered in this investigation. However, if a
significant part of the method dependent variance component is due to
errors in the manual parallax measurements, the large size of this com-
ponent, means that the precision is better than else could be expected,
especially when affine parameters are used.
Of interrest, equal to the precision, is the relation between window
size and the reliability. We could assume that as long as the geomet-
rical and radiometrical differences betweeen the images are acceptably
modelled and the areas are without too large disturbances within the
matching windows, the reliability should increase with the window size.
Geometric Window Size
Data Set Parameters 12 16 20 30 40 50
Rock No Affine 757 191 182 157 77% 72%
High1 Parameters 20% 17% 311 38% 451 517
High2 Used 3% - 67 18% - 141
Low 327 = 43% 36% = 42%
Rock Affine 627 73% 75% 807 827 821
Hight Param. used 51 97 241 337 441 507
In the table above, showing reliability, we could see that when affine
parameters not are used the window size 50X50 pixels could be close to
the window size that gives the overall best reliability. When affine
parameters are used, the reliability inreases with increasing size of
the windows for all the tested window sizes. It is obvious that large
window sizes (>30X30 pixels) are needed in order to make optimum use of
the affine parameters, both with respect to precision and reliability.
The investigation also showed that the minimum root mean square was
reached at larger window sizes in the y- than in the x-direction both
with and without affine parameters.
GROSS ERROR DETECTION
A distinction has been made between precision, the root mean square
+583
