ified and finer- 
removed. The 
he current level 
| by a convolu- 
of successively 
ional) Gaussian 
les is by means 
quation 
ages a discrete 
» 1994). 
gh to carry out 
s is not the case 
0st importance 
coarse scale no 
lental structure 
Is this criterion. 
erg (1994) has 
space features. 
hey are smooth 
round and stand 
e four different 
lation, merge, 
tion are not too 
| 1994)), merge 
only one means 
Tore commonly 
ck and Shapiro 
ents which can 
elimiting edges 
Yolded versions 
moothed (from 
on (upper left) 
WO regions are 
1 (lower right). 
ed. Imagine a 
ted by a small 
It in two edges 
thing only one 
the term scale- 
per referring to 
1s for ideal im- 
| the road, The 
rom top to bot- 
s. The level of 
vel of noise in 
ects. Linked to 
le (Koenderink 
window which 
ale is the scale 
ir. For instance 
es in the upper 
le corresponds 
tion). 
    
  
   
  
  
  
   
   
Figure 1: Events in scale-space — a) (left) Images b) (right) Thresholded images (from left to right from top to bottom: original image, 
t = 2,t = 5: split into two regions, t = 20: merge into one big region) 
  
Figure 2: Car on road in scale-space — Images (from left to right, 
from top to bottom: original image, t — 2, t — 3,1 — 5) 
WHAT IT 
+02 -0-—-I4 
mago-dozEmmco 
Figure 3: H — This is not what it seems to be 
3 LINKING ABSTRACTION AND SCALE-SPACE 
EVENTS 
One ofthe interesting properties of the human visual systemis that 
it appears to represent information on multiple scales (Kosslyn 
1994). Figure 3 shows information on two different levels of 
scale (similar patterns are used in psychological experiments). 
On a coarse scale you can see the letter "H", but on a small scale 
this is questioned ("this is not what it seems to be”). In this 
example the information on the two scales is independent but in 
many cases there is a close interaction of large scale and small 
scale. 
How abstraction can occur by means of change of scale, how 
this is linked to scale-space events, and how this can be used 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
in practical applications is shown in the following using the ex- 
traction of roads from aerial imagery and the generalization of 
buildings as examples. 
3.1 Extraction of Roads from Aerial Imagery 
There is a lot of work in the area of the extraction of roads 
from aerial and satellite imagery (Airault et al. 1994, Gruen et 
al. 1995, Jedynak and Rozé 1995, McKeown Jr. and Denlinger 
1988, Steger et al. 1995). But only Steger et al. (1995) use 
different scales. This paper is in the line of Steger et al. (1995), 
but tries to add a theoretical foundation. 
3.1.1 Large Scale: Roads in images of a resolution of 0.1m 
up to 1m pixel size, i.e. large scale, are complex objects (see 
Fig. 4 for a simplified semantic network representation). Roads, 
or more specific the road segments, are composed of the pavement 
(elongated region made of concrete or asphalt) with colored mark- 
ings which are parallel to each other. Additionally, cars drive on 
roads, trees or buildings cast shadows on them or occlude them, 
and sidewalks lead parallel to them. 
From this follows that roads can be detected either based on 
parallel edges which are the boundaries of the homogeneous elon- 
gated regions made of concrete or asphalt and the surroundings, 
or by directly detecting the regions by means of region growing. 
Figure 5 shows the result of region growing after selecting bright 
regions, eliminating small regions and closing of holes (see Fig. 5 
b) for the original image). Even when using large thresholds to 
eliminate disturbances, as has been done here, parts of the roads 
where trees are casting shadows were not connected. A centerline 
representing the centers of the roads can be computed by simple 
thinning of the regions (Fig. 5 b)). 
3.1.2 Small scale: In images of a lower resolution (2m up to 
10m or more), i.e. smaller scale, cars on a road are not visible any 
more. The road segments are changed into linear objects which 
construct the road network (see Fig. 6 for a simplified semantic 
network representation). The road segments are connected by 
the crossroads. Special types of crossroads are simple crossroads 
or intersections. A specialization of a road segment is a bridge 
which crosses other objects. 
Because a lot of detail is missing, the model represented by 
the semantic network has a higher degree of abstraction than 
that of the large scale. Roads appear as lines. Because of their 
525