because additional 
ar reflections. 
y a metallic ruler 
sts plinth and as 
model axis served 
‘easons, the anlog 
nt equipments. 
  
ibution 
ore produced with a 
ata from the Rollei 
an processes were 
tribution of the gray 
bution for a picture, 
the objects surface 
ion was based on 
jerformed at an 
due to the lack of 
le variations of the 
iodels resulting in 
lual object points. A 
igital CCD camera, 
storing the image of 
jating mark. For all 
me object point the 
screen allowing to 
> actual image with 
pt speeded up the 
educed the amount 
leasurements were 
jerformed with the 
p with an internal 
age measurements 
0.15 mm for the 
results show a good 
juration and are the 
ous DOM out of the 
ECT MODELS 
h the object surface 
em (X,Y,Z), the use 
   
of a standard image matching procedure makes it 
necessary to define local coordinate systems 
(XmiYmi-Zmi: i=1,nmodel) for the individual stereo 
models (cf fig. 5). This is due to two reasons: 
> standard matching processes describe surfaces with 
Z(X,Y) functions, thus driving the point selection 
process from the XY-coordinates 
— the mean camera axis of adjacent stereo models have 
a longitudinal tilt resulting in a corresponding 
inclination of the XY-planes of these models. This 
inclination avoids the transfer of the distribution of 
object points chosen for the image matching from one 
model into the adjacent one. 
The Z axis of the local model coordinate systems are 
chosen parallel to the mean camera axis, what means, 
that the local XY-plane approximates the tangential plane 
of the object surface. 
The use of such a model coordinate systems makes it 
necessary to perform a transformation step of orientation 
and point data from the unique X,Y,Z system into each 
individual model system. 
  
  
  
Figure 5: Global, surface ()s and model ()m coordinate 
systems 
Image matching. As matching tool the program ARCOS 
has been used. The program is founded on an area 
based matching strategy keyed to the determination of 
more or less steady object surfaces, which will be 
described by a dense grid of regular distributed points. 
Practical tests have shown (Bennat, 1990; Gülch, 1994), 
that the program produces very accurate results even in 
cases of low image contrast. Just the latter aspect is of 
importance here due to the shape of the given object. 
As already mentioned, the image scale is about 1:8 with a 
base to height ratio of approximately 1:8. According to 
these values a parallaxe of 1 pixel equals to a height 
difference of about 0.8 mm in the object space. 
Considering the height extensions of the evaluated 
surface parts, ranging from 40 up to 70 mm, maximum 
parallaxe differences of about 100 pixels have to be 
expected in a single stereo model. Although the 
maximum value won't arise between adjacent object 
points, considerable differences have to be managed 
even on short distances. Therefore the algorithm has to 
provide a large pullin range, otherwise severe 
convergence problems would occur. In addition, the 
magnitude of the parallaxe differences expresses 
  
International Archives of Photogra 
mmetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
geometric distortions forced by the surface slope, leading 
to matching failures if they are not modelled correctly. 
The calculations are based on a hierarchical strategy, 
starting with a coarse point grid, which is densified in two 
steps (point spacings: 4, 2, 1 mm). The extensions of the 
point grids ranged from 15.000 to 19.000 [mm?] or 15.000 
to 19.000 points per model. 
In order to obtain optimal results some tuning 
investigations concerning target size and matching 
threshold have been made. They showed, that 
— small targets gave a high success rate with accuracy 
problems in regions of strong parallax differences 
— great targets used together with a standard threshold 
produced problems with the success rate leading to a 
loss of accuracy in regions of strong parallax 
differences 
— great targets used together with a lowered threshold, 
dynamically adapted to the image contrast gave the 
expected success rate (98 %) and accuracy 
The behaviour can be explained by the interrelation of 
the low image contrast and the influence of geometric 
distortions. 
= Small targets have low information content and in 
case of low contrast does this lead to statistical 
similarities although the geometry might be modelled 
incorrectly. Consequently the surface can not be 
traced succesfully, what is especially 
disadvantageous for steep surface slopes. 
= Great targets are more influenced by geometric 
distortions. If this is not modelled completely, the 
similarity is lowered beneath the threshold, resulting in 
failures. In regions of steep slopes two or three 
successive failures then lead two inaccurate start 
values for the following matchings, what can not be 
overcome due to the low contrast information. 
  
  
  
  
  
Figure 6: Matched point grid with corresponding image 
Fig.6 shows an example for a point grid calculated. It is a 
perspective visualization of the grid, showing the front 
part of the face. Obviously, the low image contrast did not 
affect the quality of the object model, clearly reproducing 
the shape of the busts face. It simply remained the 
problem of blunders. A small number of blunders could be 
identified but not supressed. There are two reasons for 
blunders: