each line segment match is a function of the overlapped 
distance between the two. The best match is determined from 
the accumulated  constributions. Figure 4 shows the 
accumulator array, with the peak indicating the best match. 
Figure 5 shows the result of registering the model to the 
image (the model boundaries are overlaid on the image). 
Details of the approach may be found in [14]. 
Note that the described processing only provides a 
transformation that relates the models to the image. This is 
not the same as actually matching building structures in the 
model with the buildings in the image. This step requires 
much more detailed processing. We need to examine how 
many of the model features can actually be found in the image 
and whether they are sufficient to confidently predict the 
presence of the building. Details of such processing are also 
given in [2]. 
ESS 
     
Figure 1 Image from Fort Hood, Texas 
  
ec SN 
EF 
e 
c 
  
  
  
Figure 2 Model projection from expected viewpoint 
  
570 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
3.3 Unknown Camera Orientations 
     
If several parameters of the camera orientation are not 
known, the process of finding the best transformation as 
described above becomes much more complex. Instead of 
searching for two unknowns, we may need to search for five 
or more unknowns. In principle, the search can still be 
conducted as above, by  hypothesizing different 
transformation parameters and computing a match score, but 
search in a five dimensional space may become prohibitive. 
An alternative is to use alignment [11] techniques. Here, a 
transformation (alignment) is computed from a small number 
of feature matches; the transformation can then be used to 
verify matching of the remaining features. The minimum 
number of features needed to estimate the transformation 
depends on the nature of features (points or lines, 2-D or 3-D) 
and the complexity of the estimation method. In recent work, 
several methods have been developed that provide closed 
form solutions for computing the transformation from the. 
matched features. | 
The alignment approach avoids searching through the 
transformation space. However, it requires correct matching 
of initial features used to estimate the transformation. If only 
low-level features, such as points or lines are used, it may not 
be possible to obtain unambiguous matches. One commonly 
used approach is to match all subsets of features in the model 
to all subsets of features derived from the image. Clearly, 
such computation can be very expensive (it is O(n™),where n 
is the total number of features and m is the number required 
to compute a transform). Use of higher level features, either 
groups of features or even better, surfaces and volumes, can 
greatly reduce the complexity by reducing the ambiguity. 
Groups of features have been used for estimating the pose of 
objects in indoor scenes (for example, see [13]) application in 
outdoor scenes is likely to be much more difficult. 
4. MATCHING FOR DEPTH ESTIMATION 
To extract 3-D structure from two or more images, we need 
to compute correspondences between points or features in the 
multiple images; good surveys of various approaches can be 
found in ([5],[6], [8]). One fundamental difference between 
this task and that of registering an image to a site model is that 
a single, global transformation is not applicable. Rather, the 
transformation from one image point to another is a function 
of the unknown height of the point. Thus, we can not use 
global matching methods but need to match points and 
features in smaller areas. Small area matches, on the other 
hand, are highly ambiguous. A pixel by itself can only be 
characterized by an intensity value; this value varies 
somewhat with the viewpoint and many pixels in an image 
will have similar intensities. Some context from the 
neighborhood needs to be utilized to disambiguate.