(26) 
ce 
0 
og 
(27) 
iability 
AP ba- 
mizing 
> min, 
(28) 
he sur- 
lent on 
ould be 
the in- 
' js so- 
t good 
n of T. 
'ge our 
version 
on pro- 
' or en- 
>. This 
| inver- 
ve can, 
Iterati- 
robust 
against 
liabili- 
Hahn, 
ds for 
orithm 
on (cf. 
ta sets 
    
   
< WW 
DOSED 
oco oS 
SELLES 
<> do eS OC Seo 
      
  
      
4 
% 
% 
d 
5 
i 
Y 
On 
KX 
SS 
A 
A, 
$ 
5 
M 
() 
i^ 
^ 
d? 
& 
      
«S 
S 
o 
N 
5 
y 
A 
X 
M 
M 
Oo 
X 
0 
gas 
X 
% 
QA) 
e 
520 
=) 
  
  
OO 
LOX 
a) 
A 
0 
^ 
  
   
K 
  
    
  
000 
X 
X 
    
Q 
   
      
  
() 
% 
Q 
X 
M 
KS 
(> 
NO 
( I 
c] 
NS 
=) Xi 
     
   
  
    
Figure 4: The computed surface and its a posteriori 
accuracy 
Let us firat look at a stereo pair of digital aerial 
images shown in Figure 3. They represent a piece of 
steep and rough wilderness with rock-debris. Each of 
them has 240 x 240 pixels. The image scale is about 
1 : 10000. This image material was also used to test 
the feature based and least squares matching algo- 
rithms and is regarded as the hardest one within three 
selected projects (cf. Hahn/Fórstner, 1988). 
Figure 4 shows the automaticly generated surface field 
and its posteriori accuracy. It contains 30 x 30 lat- 
tice points with 1 x 1 m? lattice size. All surface 
heights of the same lattice points (900 points) was 
also manually measured on an analytical measuring 
device Planicomp C 100 as reference (cf. Fig. 5). 
The precision of the manual measurements is about 
  
495 
HL 
  
Figure 5: The same surface measured manually 
MEAN: —O.313 m SDEV: 0.207 m 
  
  
  
  
  
o 200 400 600 800 
Figure 6: The difference between two surfaces 
0.22 m (x 0.14 °/,, of flying height). Figure 6 illustra- 
tes the difference between the automatically and the 
manually generated surface fields. This difference can 
be characterized by its mean (bias) and its standard 
deviation against the bias. Taking the a posteriori ac- 
curacy of the automatically generated surface (cf. Fig. 
4) into account, the results are: 
MEAN DIFF: —0.313 m (bias), 
SDEV: 0.207 m (2 0.13 ?/,, of flying height), 
where the precision of the surface reconstruction using 
our algorithm is about the same as observed by the 
operator. 
Finally, we look at the image pair “House” (cf. Fi- 
gure 7), which is one of the twelve image pairs for the 
test on image matching of the working group III/4 
of International Society for Photogrammetry and Re- 
mote Sensing (cf. Gülch, 1988). This image pair has 
been classified by the test organizer into the group 
of high complexity for image matching, as it contains 
almost all troubles, including discontinuities, occlu- 
sions, shadows, and corruptions. Each image has a