How the Precision was Measured
In this investigation I have tried to make a distinction between the
reliability and the precision of the result. How could the precision be
measured? The differences in precision between the methods have been
computed using the same check points for all methods which are directly
compared with each other. The points have been manually chosen so as
not to have gross errors in any of the results. The precision measure
used is the root mean square (r.m.s.) deviation between the manual mea-
surements and the matchings for the chosen points. The new sub sets of
results thus obtained are noted with the name of the data set followed
by the number of used points within paranthesis (example, "Rock(30)").
How the Reliability was Measured
In this investigation the term reliability is used as a measure of the
lack of gross errors. Reliability was measured as the averages of the
points with radial errors lower than 15, 20 and 25 ym or 10, 15 and 20
um for errors in the x-parallax for one-dimensional methods.
Do Radiometric Parameters Improve the Accuracy?
A question to be answered is if simultaneous matching and computation
of radiometric parameters have advantages in comparision with making
the normalisation prior to the matching? No significant increase in
precision was obtained by implementing radiometric parameters as un-
knowns while a small positive effect on the reliability was observed.
The multiplicative parameter was more important than the additive, pro-
bably because of the better correspondence to differences in exposure.
Do Geometric Parameters Improve the Accuracy?
In order to reduce the negative influence from the geometrical differ-
ences between the images, additional geometric parameters could be
used. The parameters should model unknown projective differences bet-
ween the images. As the windows are small, 4 affin parameters could be
used when the rotations not are known, or alternatively, 2 linear para-
meters in the stereo direction when the images are relatively oriented
and the rotation is known. In the table below, the root mean square
deviation in um is shown for some matchings with (AP) and without (No)
affine parameters, for some different window sizes.
Window Size
Data Set 12 X 12 16 X 16 20 X 20 30 X 30 40 X 40 50 X 50
No AP No AP No AP No AP No AP No AP
Rock{(30) (10.7 10.3} 9.5 9.5 9,9 9,1 11077 6.5 110.5 8.5 13.0 975
Low(15) - = = - 10.1:10.3 - - - - - -
The increase in precision got when using affin parameters seems to be
of practical importance for window sizes above 30X30 pixels. For the
window sizes 30X30 and 40X40 pixels the hypothesis that no precision
change is got when using affin parameters are rejected on the 5% level
with the Wilcoxon test, for the hypothesis that affin parameters have a
- 580 -
