Jorge Brito 
  
  
    
  
      
It is worth mentioning that these statistics mentioned above were 
computed for both images simultaneously. If the comparisons were 
issued for individual images, the results would be even more 
encouraging. See, for example, the results depicted in table 2: 
    
4, 6, 10, 11, 17, 
18, 21, 24, 25 
SIMULATED 
STEREO 
  
  
     
1, 2, 3, 7, 8, 9, 
  
    
  
  
  
12, 13, 14, 15, 
i 2 2, Image STEREO vs. ORTHO SIMULATED vs. 
: ORTHO 
110 (16/26) =61.5% (18/26) = 69.2 % 
112 (22/26) = 84.6% (22/26) = 84.6% 
  
4, 6, 10, 11, 17, 18, 
21, 24, 25 ; : : 
Table 2. Summary of the comparisons between single images 
ORTHO 
Figure 8. Venn Diagram showing the 
simultaneous analysis of occlusions for the 
26 point pairs measured in the stereo-model, 
simulated data, and orthoimage. 
5 CONCLUSIONS 
The results discussed above clearly indicate the feasibility of using a stochastic method (the Bayesian network) to deal 
with the problem of automatic detection of occlusions in digital images. 
The following conclusions can also be drawn from the tests implemented with those data sets: 
(1) identification of the causes of the occlusions can be solved by elimination. Such elimination depends on available 
knowledge about the images. For example, if a pair of points is known to be in an occluded area, but if it cannot be 
detected through the model, then the hypothesis of relief displacement (geometric disagreement) can be excluded. In 
that case, the occlusion can be attributed to shadows of buildings or clouds in the images; 
(2) digital orthoimages can be used for occlusion detection analysis. 
The most important conclusion, however, is the relationship between the analysis of occlusion performed by the 
implementation and the terrain itself. In other words, the terrain portrayed in the images “totally agrees” with the results 
of the occlusion analysis given by the implementation. 
REFERENCES 
Brito, J.L.N.S., 1997. Precision of Digital Orthoimages: Assessment and Application to the Occlusion Detection 
Problem. Doctoral Dissertation. The Ohio State University. Department of Geodetic Science and Surveying. Columbus, 
Ohio. 
Charniak, E. 1991. Bayesian Networks Without Tears. In AI Magazine. Volume 12. Number 4. Pages 50 - 63. 
Doorn, B. D., 1991. Multi-Scale Surface Reconstruction In the Object Space. Report Nr. 413. The Ohio State 
University. Department of Geodetic Science and Surveying. Columbus. Ohio. 
Li, D., Shugen Wang, and Rongxing Li. 1996. Automatic Quality Diagnosis in DTM Generation by Digital Image 
Matching Techniques. In Geomatica, Volume 50. Number 1. Pages 65 - 73. 
Norvelle, F. Raye. 1992. Using Iterative Orthophoto Refinements to Correct Digital Elevation Models (DEM's) . In The 
ASPRS/ACSM 92 Technical Papers. Volume 2. Pages 347 - 355. 
Pearl, J., 1988. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kauffmam 
Publishers Inc.. 
Stassopoulou, A. 1996. Bayesian Networks for Inference with Geographic Information Systems. Department of 
Electronic and Electrical Engineering. University of Surrey. Guilford, U.K. 
  
108 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.