Figure 7: The image pair “House” 
size of 240 x 240 pixels and the image scale is about 
1 : 3000. In Figure 8, we show the computed surface 
field, by means of a perspective view and a contour 
map. 
9 Conclusion 
In this work, we concentrate our attention on in- 
ference processes in computer vision and formulate 
many of its goals in a general manner as a ill-posed 
problem of image inversion. Based on MAP criteria, 
we have introduced a theoretical framework for dea- 
ling with ill-posed inverse problems. We have shown 
how the surface reconstruction from images can be 
solved under this theoretical basis as an application. 
Of course, this is only a limited domain of its app- 
lications. So, among the goals of future work will be 
1) the introduction of a learning mechanism to im- 
prove and adapt the a priori knowledge during the 
inverse process, 2) the application of neural network 
technology to developing parallel algorithms for sol- 
ving optimizing problem mentioned above, and 3) the 
extension of the application range of the approach. 
References 
[1] Chou, P. B. and C. M. Brown,1988. Multimo- 
dal reconstruction and segmentation with Mar- 
kov random fields and HCF optimization, Proc. 
Image Understanding Workshop, San Mateo, 
CA: Morgan Kaufmann, pp. 214-221. 
[2] Fawcett, J. A.,1985. Inversion of N-dimensional 
spherical averages, SIAM J. Appl. Math. 45, pp. 
336-341. 
[3] Geman, S. and D. Geman, 1984. Stochastic Rela- 
xation, Gibbs Distribution and Bayesian Resto- 
ration of images, IEEE PAMI, Vol. 6, pp. 721- 
741. 
[4] Giilch, E., 1988. Results of Test on Image 
Matching of ISPRS WG III/4, Intern. Archives 
of Photogrammetry and Remote Sensing, vol. 27, 
Kyoto. 
496 
SEES 
e 
     
    
Z Cm) 
RAIRRIGEGES 
ES 
4 $P 
  
  
  
  
' 
z 
72 
4 64.75 
dL LLL dort / LLL 
64.00 64.00 
Gin L 
14.00 14.75 15.50 16.25 17.00 17.75 18.50 19.25 20.00 20.75 21.50 22.25 23.00 23.75 
SCALE 1:1 473 
LT ERE 
Figure 8: The computed surface 
[5] Hahn, M. and W. Fôrstner, 1988. On the Appli- 
cability of a Feature Based and a Least Squares 
Matching Algorithm for DEM Acquisition, In- 
tern. Archives of Photogrammetry and Remote 
Sensing, 27-3, Kyoto. 
[6] Poggio, T., V. Torre, and C. Koch, 1985. Compu- 
tational Vision And Regularization Theory, Na- 
ture, vol. 317, pp. 314-319. 
[7] Tarantola, A.,1987. Inverse Problem Theory, El- 
sevier, New York. 
[8] Zheng, Y.-J., 1990. Inverse und schlecht gestellte 
Probleme in der photogrammetrischen Ob- 
jekt Rekonstruktion, /naugural- Dissertation, Fa- 
kultät für Vermessungswesen, Universität Stutt- 
gart, Stuttgart. 
[9] Zheng, Y.-J. and M. Hahn, 1990. Surface Recon- 
struction from Images in the Presence of Discon- 
tinuities, Occlusions and Deformations, Intern. 
Archives of Photogrammetry and Remote Sen- 
sing, vol. 28, part 3/2, pp. 1121-1144, Wuhan. 
[10] Zheng, Y.-J., 1992. Inductive Inference and In- 
verse Problems in Computer Vision, Robust 
Computer Vision, Fôrstner/Ruwiedel (Eds.), 
Karlsruhe: Wichmann, pp. 320-350. 
ABS 
Most 
used 
and 
2D + 
modi 
data 
man 
unifi 
withi 
À po 
into 
start: 
DTN 
conc 
tures 
asses 
stenc 
Key v 
tion s 
Since 
unde 
but ¢ 
terra 
boun 
terra 
Xy, tl 
one-c 
very 
struc 
node