The data used for simulation is a TIN-based three 
dimensional lines which is shown in Fig.3 , in which we 
have used shading to create a depth impression. We first 
calculate the corresponding stereo image pair by 
simulating the camera geometry. The simulated image 
pair is shown in Fig.4. 
  
   
    
  
V 
\ 
NI 
  
     
IN 
47 
D 
  
   
      
    
     
   
  
    
   
  
7 
3 
7: 
  
  
  
    
     
    
     
    
| 
X 
p 
     
    
  
  
AND 
AN 
» 
N 
IX 
    
       
UM 
Ww 
AV 
| 
2 
cz 
   
  
Fig.5 The final matching result (top view of 3-D lines) 
In the image space, as described before, we use 
overlapping, line orientation, disparity limit to find the 
candidate matching. We have 292 lines from each 
image, the result of candidate matching contains 3956 
candidate line pairs. In this experiment, we use the 
"connectivity" as our GGCS. The label value for a line 
is enhanced if it connected to the other line at the 
endpoints. The final matching result is drown in Fig.5 
(here we only give the top view of the three-dimensional 
data). The final result has 284 lines. The loss of some 
lines which are in the original TIN data is the result of 
using ordering constraint. 
Experiment on real images 
A stereo images of a stereo plotter (shown in Fig.6, on 
the last page of this paper) has been chosen for the test. 
The images were taken with a normal camera (focus 
length 28mm and size 36mm X 24mm), followed by the 
scanning with 100 DPI on positive pictures (about 3 
times larger than negative one). The 19 stereo points 
were visually identified by hands, and used to calculate 
the orientation parameters, with the result By/Bx — - 
0.00213, Bz/Bx — -0.08319, phi — -5.194648, omega 
— -0.126733, kapa — 0.247479 (phi, omega and kapa 
are in degree). 
The program started with sobel operator which produces 
image gradient magnitude and orientation, later filtered 
by Duncan's technique. Using the methods described in 
section 2, the resulted line detection is illustrated in 
Fig. 7, The short line segments were sorted out in order 
to reduce the number of candidate matching. The 
number of image lines involved in the matching is 83 on 
both images. The final matching has 55 lines, which has 
turned out to be a correct matching. 
538 
8. CONCLUDING REMARKS 
This paper has presented a new approach to solve the 
correspondence problem. The heard of the approach is 
the back-projection of image space based candidate 
matching result into the object space and solving the 
stereo matching as a consistent labelling problem. The 
experiment has demonstrated that it is a promising idea 
to tackle this difficult problem. Also, our research has 
shown the remaining difficulties in robust line detection 
or extraction, choice of matching attributes and 
thresholds, etc. The future research will be oriented in 
dealing with these problems, author wish that more 
results would come up soon. 
9. ACENOWLEDGEMENT 
This research is jointly supported by CCGM (Centre for 
Computer Graphics and Mapping) and FRANK project, 
Faculty of Geodesy, TU Delft. FRANK is a registered 
trademark of FRANK system supported by Geeris 
Holding Netherlands BV. The computational support 
from Photogrammetry Laboratory is also appreciated. 
10. REFERENCES 
1. Wrobel, B., "Facet Stereo Vision (FAST vision) - a 
new approach to computer stereo vision and to digital 
photogrammetry", In proceeding of intercommission 
conference on Fast Processing of Photogrammetric 
Data, Interlaken, Switzerland, pp.231-258, 1987. 
2. Helava, U.V., "Object-space least-square 
correlation”, Photogrammetric Engineering and Remote 
Sensing, Vol.54, No.6, Partl, June 1988, pp.711-714. 
3. March, Riccardo, "A regularization model for stereo 
vision with controlled continuity", Pattern recognition 
letters 10(1989) pp.259-263. 
4. Ackermann, F. and Zheng, Y.-J., "Inverse and ill- 
posed problems in  phtogrammetric surface 
reconstruction", In proceeding of ISPRS Commission III 
conference, Wuhan, China, 1990. 
5. Zheng, Y.-J and Hahn, M., "Surface reconstruction 
from digital images in the presence of discontinuities, 
occlusions and deformations", In proceeding of ISPRS 
Commission II conference, Wuhan, China, 1990. 
6. Baker, H. Harlyn., " Surfaces from Mono and Stereo 
Images ", Photogrammetria, 39 (1984) pp.217-237. 
7. Benard, M., "Automatic Stereophotogrammetry: A 
method Based on Feature Detection and Dynamic 
Programming ", Photogrammetria, 39 (1984) pp.169- 
181. 
8. Greenfeld, J.S., " Experiments with Edge-Based 
Stereo Matching ", Photogrammetric Engineering and