by the image densities in the neighborhood of the inter- 
polated surface element. 
At the end of the rectification process at least two sets 
(one for each image) of albedo values are available 
surrounding each grid point. For every grid point these 
sets are then correlated. As usual the location of the 
highest similarity value within a 3 by 3 vicinity is calcula- 
ted. Subsequently, the object height has to be compu- 
ted from the resulting parallax. 
Finally, the residuals between the a priori heights and 
the new ones are calculated and combined to a mean 
square value which is checked against a threshold defi- 
ning the success of the computations. 
  
     
     
surface element 
Been wine = 
W = grid point 
Fig.3 Projection of the image densities onto the surface 
Control aspects À flexible and purely automatically 
operating algorithm needs some internal control to 
assure satisfying functionality. Out of the numerous 
features three will be mentioned here. 
First, the appearance of failures has to be taken into 
account. Especially due to the limitation of the window 
size to a fix value, defined by the grid width and the size 
of the surface elements, in cases of low contrast failu- 
res will be as numereous as with extendable windows. 
This disadvantage might be compensatable, because all 
points belonging to a homogeneous structed surface 
part are calculated in parallel. This allows the substituti- 
on of failed calculations by computations from the 
parametric surface description. Although these substitu- 
tes will be reliable, the points affected have to be 
marked to allow an a posterori assessment of the re- 
sults. 
Secondly, we have to consider those determinations, 
which will result in blunders. Manifold reasons might be 
130 
responsible for their occurence which would not be 
identifiable immediately. However, an algorithm should 
at least perform the identification of such points. This 
might be achieved correspondingly to the substitution of 
failures. As the determination of the functional surface 
parameters in general will be a redundant process, a 
blunder detection algorithm might be applied to the 
results, in order to discover unreliable point heights. 
Finally, using a least squares adjustment for the estima- 
tion of the surface parameters permits an evaluation of 
the correctness of the functional set up. This can be 
used for internal control purposes and for selection of a 
new functional description, if necessary. 
Process tuning Besides the internal algorithmic 
control some general fixings have to be done. Here are 
the size of the surface elements and the grid widths or 
window dimensions rsp. of interest. 
The size of the surface elements defines the geometric 
resolution and therefore affects the accuray attainable. 
Although no direct interrelation exists, the selection has 
to account for the desired precision. 
Unfortunately, the tendency towards small element sizes 
will result in a reduction of the image contrast available. 
As consequence, one has to expect an increasing 
number of failures and/or blunders, thus diminishing 
the robustness of the procedure /Faugeras 1992, 
Boochs,Hartfiel 1989/. 
Furthermore, the interrelation of these two factors is 
superimposed by the image quality and the albedo varia- 
tions within the object, being responsible for the image 
contrast. The selection of the element size therefore has 
to reflect the data quality and the aims of the application. 
Regarding at the dimension of the image windows, 
defined by the grid width and the size of the surface 
elements, a similar view may be obtained. Small wind- 
ows result in higher precision /Faugeras 1992/ accom- 
panied by a loss of robustness. Although in general, 
robustness might be acquired by an adopted increase of 
the window size, this will not be necessary within the 
actual concept, because due to the contextual determin- 
ation of several points within regions of homogeneous 
shapes the calculations will be stabilised and allow the 
substitution of incalculable heights, if desired. 
Introduction and use of additional informations 
Due to the flexible concept some additional knowledge 
has to be available within the calculations to achieve 
good results. This are 
° the type of functionals to be used 
° the average size of the regions 
* the run of the region borders 
° the location of important object features (break- 
lines, edges etc.) 
These informations might be provided explicitly and in 
advance by discrete definitions or might be introduced 
internally within preprocessing algorithms, based on the 
image data available. 
The explicit definition for sure is the simpler solution, 
because there further investigations might be restricted 
to 
av 
orc 
acl 
ust 
Th 
ex 
boi 
re: 
su 
ca 
ele 
su 
ni 
We 
re 
Tt