(a) Ten dots 
(e) Proximity influence 
Figure 6: Proximity grouping 
8.2 An Example 
Consider analyzing a harbor or port complex. We wish to de- 
scribe the buildings in the port facility, the transportation net- 
work around the facilities, and of course the pier areas and the 
ships in the area. We already discussed detection of buildings 
and transportation networks. What do we need to know about 
port and harbor facilities to detect the piers and describe the 
ships? That the planning and design of port and harbor facili- 
ties is strongly dependent on the characteristics of the ships to 
be served, and the type of cargo to be handled [62]. To eventu- 
ally describe the scene completely we would have to know a lot 
of things about the ships: Main dimensions (length, beam, draft), 
cargo-carrying capacity, cargo-handling gear, types of cargo units, 
shape, hull strength and motion characteristics, mooring equip- 
ment, maneuverability, and so on. To detect only the pier areas 
(where later we would look for ships) we only need the upper 
    
   
er 
ne LL 
mts 
= 
i 
  
   
  
atat n 
    
   
E 
   
> 
cC 
   
  
   
(a) Collinear lines 
     
  
       
  
     
(b) Collinear influence 
bounds on ship dimensions and the approximate image resolu- 
tion. These parameters are easily available a priori and chiefly 
determine the extent and strength of the grouping fields associ- 
ated with the features that we will use. 
Figure 8a shows an image of a portion of the U.S. Navy facil- 
ities in San Diego. We know that we should expect to see mostly 
military ships that may require long term docking, thus allowing 
for double or triple docking. We know the image resolution and 
the approximate ship dimensions, thus we know the minimum 
size of the piers. The following steps are applied: 
Locate Boundary between Land and Water: We de- 
tect the boundary between land and water regions automatically 
using our implementation of [50]. In this example we arbitrarily 
selected the largest region to represent the water region. Next we 
approximate these boundary by piecewise linear segments, shown 
in figure 8b using LINEAR, our implementation of [48]. 
Locate *land" Structures in Water: Contrary to many 
natural structures on shore, cultural structures appear highly 
geometric. We expect that most piers will appear as linear struc- 
tures attached to the shore, and in the water. Their linearity 
indicates that the piers or portions of piers should be character- 
ized by apars (parallel groupings). Ships are typically docked 
parallel and adjacent to the piers. We then expect that most of 
the line segments corresponding to sides of piers, sides of ships, 
shadows, and so on in the neighborhood of the piers would result 
in many local parallel groupings (apars). The limit on parallel 
groupings is a function of image resolution and ship dimensions. 
The apars in our example are shown as thin lines in figure 8b. 
Detect Pier Areas: The apars are easily classified into 
“land” or “water” apar with respect to the water region. Sub- 
sequent processing operates on the land apars only. Next, we 
apply Px0D grouping to the land apars. The extent of the fields 
is task-dependent however it need to be only approximate. At 
the resolution in our example, the radii of the fields are roughly 
equivalent to the pier width plus the width of three destroyers on 
both sides of the piers, or about 16 pixels. Each apar (its center 
of mass) contributes a field. Subsequent contributions shift the 
center of mass of the group. We then select the groups so that 
apar membership is exclusive by extracting the possible groups in 
order of decreasing mass (number of apars). The resulting groups 
represent potential pier fragments (groups in figure 8b and arrows 
in figure 8c.) The representation of the resulting groups is the 
same as that of apars. 
Next we apply ColD to the pier area fragments. The longest 
piers are about three times the length of a destroyer thus we 
allow the extent of the elliptic fields (see figure 8b) to be up to 
three times the length of the apars. The width is equivalent to 
the apar width (or group radius, 16 pixels in this example). 
The result of the grouping, shown in figure 8d, is then repre- 
sented, again, by apars and denote potential pier areas. 
8.3 Saliency-ehancing operators 
The second effort deals with saliency-enhancing operators ca- 
pable of highlighting features which are considered perceptually 
relevant. These are introduced in [20]. They are able to extract 
salient curves and junctions and generate a description ranking 
these features by their likelihood of coming from the original 
         
  
    
  
i [i4 
(c) Resulting groups 
  
Figure 7: Colinearity grouping 
852