CE 
Or. 
es 
dt 
un 
re 
je 
nt 
ay 
  
  
  
161 
   
  
Wd 
Y 
NN 
   
      
    
\ 
x 
N 
RN 
NONIS 
NAT NNNM 
NAN 
“ 
EN 
SN 
M 
FFE 
al 
N 
MW 
NR 
RHR 
IN 
N 
RR 
NNNM 
RHR 
NN 
  
$500 
450 
400 
350 
0 
25 30 95 40 
n 2 
Fig. 1. Two synthetic disparity maps over the same object (data 2) as seen from two different stereo head 
positions. 
   
  
    
     
  
9) 
(/) 0» 
nl das 
(X f 
1 
lil 
% dj in 
J 
"i 
7 
iul 
Iu 
JA) 
il 
4 
IM) 
Lun 
IM 
JN Cor 
2) 
7 
i» 
y) 
/ 
nl 
ES 
) 
hy 
  
    
  
    
  
    
     
i 
À) 
^ 
'" 
IH 
i 
y 
4) 
7 
JN 
" 
  
      
[/ 
7 
lly 
hl 
J) 
wr 
ly 
JIN 
4 
i 
(/ 
Jl 
D 
        
     
J 
7 
d 
/ 
7 
0 
05 
  
Ü 
A 
4 
(n 
ZA 
        
x1 xi 
Fig. 2. An initial estimate for the relative orientation (left) and the relative orientation after surface matching 
(right) in the case of data 2. 
Table 1. Parameter estimates for the relative orientation in two cases solved by modeled features and by surface 
matching. . 
  
  
  
  
  
  
  
  
  
  
  
data method 0 Q w T. T, T. 
1 true —0.05 —0.1 0.1 —0.5 —0.1 —0.05 
1 features —0.0500 —0.1000 0.1000 —0.5001 —0.1000 —0.0500 
1 surfaces —0.0495 —0.0982 0.1030 —0.4970 —0.1027 —0.0467 
2 true —0.05 0.2 —0.02 0.8 0.01 —0.05 
2 features 0.0587 0.1809 —0.2450 0.5733 —0.1972 —0.1741 
2 surfaces —0.0498 0.1995 —0.0200 0.7993 0.0095 —0.0501 
  
  
  
  
  
  
The first data illustrates a case when the modeling succeeded. The plane normal and the axis and vertex of the 
cone were used as modeled features. The both disparity maps P3? and Q33 covered the cone better than in the 
left picture of Fig. 1 so that the modeling of the cone was successful and an accurate orientation estimation 
was obtained by the method of modeled features. The method of surface matching gave a bit worse estimation 
for the first data although the orientation obtained by modeled features was given as an initial estimate for the 
surface matching algorithm. 
'The purpose of the second data is to show how the method of surface matching works. The true orientation pa- 
rameters were thus chosen so that the conical patch in the second disparity map Q3» was tiny and consequently 
the estimated cone parameters poor. The orientation obtained by modeled features (plane normal and cone axis 
and vertex) was refined by the method of surface matching taking also advantage of the disparities in the saddle 
surface region. The refinement was considerable as seen from the figures of Table 1 and illustrated in Fig. 2. In 
Fig. 2, the synthetic maps are plotted in the X1, Y1, Zi éoordinate system corresponding to the pixels of the left 
camera in the both stereo head positions. Note that the measurement points in P32 and Q32 are not expected 
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop “From Pixels to Sequences”, Zurich, March 22-24 1995