Adopting proper price coefficients for the cost function 
is of importance to obtain reliable results. As shown in 
Eq. (13), the price coefficients form the strength of 
connections between neurons and the constant input of 
each neuron. Conceptually, one can imagine them as 
the weights can be adjusted to adapt the system to 
various conditions of applications. For example, if we 
think the factor of shape similarity is much more 
important than the other factors, we can give a large 
value to C,. However, although the concept is plausible, 
there are no rigorous rules to tune the system to be 
optimal. 
The settings of the thresholds in Eqs. (8) and (9) are 
also important. They can be adjusted to fit the 
conditions of the applied images. The threshold of 
shape similarity is set for the disturbances of image 
distortion and noises, and the thresholds of orientation 
consistency is set for the geometric changes due to the 
different view angles of cameras. 
4. DETERMINING CONJUGATE POINTS 
After conjugate features are matched, the approximate 
relative orientation can be solved by using the 
coordinates centroids of conjugate features. This 
allows us to narrow down the search windows when we 
apply template matching to determine conjugate points. 
In this application, the normalized cross-correlation 
(NCC) is used to determine conjugate points up to one- 
pixel accuracy. Then sub-pixel accuracy is reached by 
using the least-squares matching (LSM)) [Ackermann, 
1984]. 
In order to obtain reliable matches, each template 
should contain enough gray-level changes. Also evenly 
distributed conjugate points on the overlapped image 
area is required to obtain reliable relative orientation. 
    
These requirements can be achieved by using, interest 
operator to locate locally most interest points as the 
template locations. 
S. RESULTS 
Several pairs of aerial stereo photographs have been 
tested on the system. Fig. 5 shows one of the pairs of 
the test photographs. The photos were digitized in the 
resolution of 600 dpi which is corresponding, to the 
pixel size of 42.3|um square. 
By using the technique of region growing, homogenous 
areas are segmented from each image. There are 23 
features derived from the left image and 34 features 
derived from the right image (Fig. 6). 
Conjugate features are determined by using the 
technique of matching, Fourier descriptors described in 
section 2 and 3. After 5 iterations of neural network 
computation, the final status was reached and 10 pairs 
of conjugate features were discovered. The matched 
pairs are 4-4, 6-9, 8-11, 9-12, 11-15, 12-17, 19-22, 20- 
23, 21-24, and 22-28. By checking the images visually, 
we can discover that they are all correct matches. 
By using interest operator and template matching 
techniques 150 conjugate points were discovered. After 
the computation of relative orientation, 16 conjugate 
pairs were eliminated due to their residuals of y 
parallax are three times larger than the mean square 
error. Finally 20pm of the root-mean-square error of y 
parallax is obtained. The overall computation time was 
about 30 minutes when it is executed in a Pentium 90 
computer. 
  
Figure 5: An example of test image pair 
884 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996