ntion 
pace, 
ntial 
(t we 
tions 
rified 
ys to 
| fig- 
form 
iches 
arity 
epair 
al as 
The 
> 4b. 
eters 
uted 
rien- 
hese 
nd a 
Ways 
CUS- 
| are 
ve a 
eric 
Vays 
J 
= 
<< 
bor 
;ual 
ines 
sify 
f a 
this 
  
  
(c) Focus of attention 
(b) Line segments 
  
(d) Runways and markings 
Figure 4: JFK Runways detected by collinearity grouping 
9 Stereo Techniques 
The key problem with general stereo systems is the ambiguity 
in matching necessitating a mechanism to choose among many 
competing matches for each match primitive. For our system 
we have found the constraints imposed by the structure of the 
grouped features (rectangles) sufficient to select unique matches 
for the primitives. In the rare case of a rectangle having more 
than one match, we choose the match with the least number 
of disparity differences between the sides, which is equivalent to 
preferring the least occluded interpretation. In the following we 
summarize from the work of Mohan and Nevatia. Full details 
can be found in [44, 47]. 
849 
The rectangles detected monocularly as described in the pre- 
vious section are used in this system for stereo matching, object 
segmentation and shape description. Stereo matching is per- 
formed on the rectangles to obtain height information. Strue- 
tural reasoning is performed on the matched rectangles, based 
on monocular information and 3-D information obtained from 
stereo, to segment building roofs. The object segmentation au- 
tomatically provides a shape description of the roof in terms of 
the component rectangles. The segmented roof area and their 
heights are used to generate 3-D models of the buildings.