Sidewalk ~ . parallel 
- 
- 
- 
  
  
  
  
— —»- Part-of relation 
a crm 3 Road segment | ----— General relation of objects 
drives on .--* 
  
  
=" casts fd. 
| Tree | shadow | Pavement | | Marking | ; parallel 
  
Figure 4: Model for large scale (concepts are depicted in boxes) 
  
x 
à p. 
Re 
  
Figure 5: Region growing — a) (left) Large bright regions, b (right) Skeletons on top of the original image 
material (concrete or asphalt), they mostly appear brighter than 
the background. 
In Figure 7 hypotheses for centerlines of roads (bright lines) 
created in images of different resolutions (image pyramid; sim- 
ilar to scale-space but non-linear effects arise by subsampling) 
are shown. Lines are found using non-linear search, tracking 
and global non-maximum suppression with fixed parameters. 
Whereas in the image of larger scale (Fig. 7 a)) only some broken 
centerlines of the road from left to right have been found, the 
centerline was found quite stable in the image of smaller scale 
(Fig. 7 b); missing topological connections and the effects at the 
margin of the image are due to the implementation). 
In the image of small scale it was evidently possible to detect 
the whole road network. Whereas in Fig. 7 a) the effects of the 
shadows have not been smoothed away sufficiently, they do not 
disturb the recognition of the lines any further (Fig. 7 b)). 
3.1.3 Model: If one compares the large and the small scale, 
it is evident that information has been lost by eliminating regions 
and edges. This results in the elimination of noise as well as of 
substructure of objects. In Figure 8 the knowledge about road 
segments in the two scales is put together into one model. The 
model is split into three levels. The real world level consists of 
the objects and their relations on a natural language level. In the 
large scale a road is constructed of a pavement and the markings 
(solid or dashed) and cars drive or park on it. 
The objects in the real world level are connected to the objects 
in the geometry and material level by means of the concrete 
relation (one of the other relation of section 2.1) which connects 
concepts describing the same object on different levels, i.e. from 
different points of view. The geometry and material level is an 
intermediate level which represents the three-dimensional shapes 
of objects as well as their material. This level has the advantage 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
that it represents in contrastto the image level objects independent 
of sensor characteristics. 
In small scale road segments are linked to mostly straight bright 
lines in the image level via the mostly straight concrete or asphalt 
lines in the geometry and material level. 
In contrast to this the pavement of the large scale is linked to the 
elongated bright area in the image level by way of the elongated 
flat concrete of asphalt area in the geometry and material level. 
The markings are related to bright lines via colored lines and the 
car is a dark region as the concrete of a painted box made of glass 
and metal. Conceptually two things have happened here: 
e The type of the geometrical representation has changed. The 
elongated area has been condensed into a line. 
e But more important, the substructure of objects in the small 
scale (the markings, or the car on the road), has been elimi- 
nated. 
This means that the complex object road segment composed of 
region-like pavement, and with markings and cars on itis changed 
into a more abstract linear road segment. Abstraction has occurred 
by means of elimination of substructure (annihilation) and merge 
of different parts separated by effects of noise. 
3.2 Generalization of buildings 
An area where multi-scale representations are closely linked to 
levels of abstraction are GIS and cartography. Whereas a map of 
very small scale contains only the largest countries and islands, a 
large scale map shows details of buildings. It is interesting that in 
maps the outer scale typically scales in proportion with the inner 
scale (Lindeberg 1994). This means that the content is adapted 
526 
     
   
   
    
    
   
   
    
    
   
  
    
    
    
   
   
  
    
     
    
    
   
   
    
    
   
    
to the | 
abilitie: 
If di; 
Figure 
and sim 
of scal 
of the « 
betwee 
splittin 
are me 
Summi 
image 
smalle; 
e T 
scale €