Approximate values of unknowns H0, H1 are taken from neighbouring
points, AO is approximated using the parabolic interpolation in 3).
In addition to precise determination of parallax and slope an estimate
of inner precision is also obtained. This could be used to good
effect in a subsequent interpolation in a regular grid. In the
present work the inner precision is used to evaluate the interest
operator.
3. EVALUATION OF INTEREST OPERATORS
The interest operators were tested with a stereo model of a mainly
agricultural area with few buildings and only about 5% forest. Ground
topography was such that there were very few occluded areas. Photo
scale was 1:12500 with an overlap of 80%. An OPTRONICS digitizer was
used to generate 50-micron pixels.
In order to fully test the effects of each operator a relatively large
search area was set around each grid point, thereby providing a choice
of features. The window size used by the operators was generally 21 x
21 pixels. The same window size was also used in the subsequent
matching process.
To select suitable features, the following interest functions were
applied within the window of the interest operator and are detailed
below:
1. comparison of sums and variances of grey levels;
2. summation of grey level differences and second differences;
3. summation of products of grey levels.
The result of the function is called the interest value (IV).
After applying the search process to find the feature in photo 2,
standard deviations (SD) of the unknowns and cross correlations (CC)
are obtained. Additionally, the dependence (R) between SD and IV and
between CC and IV is computed. Since the interest value is not
normally distributed, R is determined from rank correlations according
to Spearman (see also /11/):
R=1-6 35, / (n3-n) , where n = number of points
i 500 for this test
D.
pK) - K,(SD) resp. D; = K,(IV) - K (CC) , 1 = 1... 0
I
K is the rank as defined by Spearman. It can be shown that an R-value
less than 0.16 has no significance (level of significance 0.1 %). In
the tables which follow, SD, CC, and R are presented against SD- and
CC-values for the original raster points.
3.1 Comparison of sums
The intention here is to find edges in the y-direction. Two
subsections of the window are always compared to give the interest
value IV.
- 567 -