232 
colour IR image. Taking the I or the Y components of the colour transformed images gives results close to 
the quality of the green channel. The estimates of noise show an inverse behaviour compared to the precision 
estimates but with a considerably smaller amplitude. The stronger texture i.e. the larger edge busyness is 
obviously correlated to the estimated noise level. The conclusion of these results is that under the given cir- 
cumstances (points selected by the operator, etc.) for matching colour images the simultaneous multichannel 
solution should be computed or the red (or IR) component of the colour image should be used. 
3.4 Empirical precision 
The procedure for assessing precision of colour matching empirically is outlined in section 3.2. The points 
selected interactively by an operator in one image have to be matched in the second image of the image pair. 
The estimated location of that point in the second image is compared to the operator controlled measurement 
of the conjugate points. As in the case of the theoretical precision the sample size is 122 points in the Glandorf 
project and 120 points in the Schorndorf image pair. In addition, a relative orientation with the matched points 
is computed. Again the series of different window sizes mentioned above for the Glandorf project and the 
Schorndorf project is used. 
First we check for systematic differences between the automatic and the operator controlled measurement. 
The mean difference between the location of the matched points and the interactively (at higher resolution) 
measured position is determined. The differences in the Schorndorf project are plotted in figure 2 (e) and (f). 
Even though they are smaller than 1/10 of the pixel size the differences in the x-direction have a certain trend 
(figure 2(e)). In the worst case of the Glandorf project a difference of about 1/10 th of the pixel size is observed. 
This difference occurred in the x-direction of matching the blue channels at a window size of 517 pixels. The 
data did not show any significant systematic difference. Therefore, in the following rms values between manually 
and automatically measured stereo point locations of the given points are calculated for the assessment of the 
empirical precision. 
The precision of the different solutions of matching the colour image pair are plotted in figure 1 for the Glandorf 
project and figure 2 for the Schorndorf project. The first two rows in both figures show the rms values for the 
location of the matched points in x and y direction. In the third row of figure 1 the precision estimated by 
relative orientation is plotted. Even on the first view the similarity between the plots of the second and third 
row can be recognized. This means that y parallaxes calculated by relative orientation and the rms differences 
in y are fairly close. Nearly the same situation occurs in the Schorndorf project. 
The values along the horizontal axis of figures 1 and 2 are the width of the quadratic matching windows. The 
plots in the left column are the results of matching each channel R, G, B of the colour image separately. In the 
right column the multichannel solution and the results for the I and Y component are visualized. 
The rms values of matching the blue component show a smaller accuracy than the rms values of all other 
solutions. The green and the red component lead to similar rms-values with a certain advantage for red. This is 
quite conform to what is found by the theoretical precision. The increasing rms-values for very small windows 
(15? pixels and below) could be expected. Furthermore, the decreasing precision for large windows (45? pixels 
and more) is no surprise. À main reason for that behaviour are disturbing 3D objects but also the insufficient 
geometric modelling by the affine transformation. 
Looking to the right column of figure 1 shows practically the same precision for the multichannel and the I and 
Y channel results of image matching. This means that for the Glandorf project the gradation indicated by the 
theoretical precision values is not confirmed. A main reason could be that in most of the selected points all 
colours are present. Because in the RGB representation intensity information is included in all three components 
we suppose that a very high correlation of the R, G, and B components leads to this result. Furthermore, the 
multichannel result is close to the quality of single channel matching with the red and green component. 
The differences computed with the IR colour images of the Schorndorf project (figure 2) are larger than the 
ones derived from the RGB colour image pair. Mainly for the windows smaller than 31? pixels the mc results 
are more precise than the I and Y results. Again as in the case of the RGB images the quality difference to red 
(IR) and green (R) is not very dominant. 
3.5 Convergence radius 
Finally some experiments are carried out dealing with the convergence radius of matching. The point location 
of the conjugate points measured interactively are used as starting points of matching. To find out the limit 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995