an be coded in a 
ales. 
s segmentation for 
ised on a masked 
elds and of “don’t 
n was originally 
:d characters. The 
ith blurred edges. 
to search for the 
vindow. We place 
ields, which helps 
on coefficient c is 
3:) (1) 
"off" mask, g. - 
covered by “on” 
he image covered 
in the mask, m— 
nber of “on” and 
iation of intensity 
lard deviation of 
ipper right, lower 
d with the whole 
coefficient higher 
selected pixels are 
rimitive instances 
| "lower left") and 
exemplary corner 
blue and yellow 
than the detection 
bute. For a typical 
in from texture of 
  
ide textures: red - 
ower right" and 
a human observer 
g Marr's principle 
would be entered 
sing an according 
ce. However, this 
  
  
overloads the computational resources necessary for the 
following reasoning currently available. So we decided to 
perform a non-maximum suppression as is usually performed in 
computer vision. Only 365 L-primitive instances remain which 
are displayed on black ground in Fig. 3. 
Figure 3. Primitives extracted after non-maximum suppression 
From this figure the reader may estimate what is lost during the 
primitive extraction phase. Recall that the following symbolic 
process only sees these data, not the image itself. 
22 Two Different Production Systems 
The paper reports experiments with different types of 
production rules representing lattice grouping and symmetry 
grouping. Three variants are compared — see Tab. 1: 
1) “canonical” the natural common sense part-off hierarchy is: 
A facade consists of a vertical column of (two or three) 
horizontal rows of (e.g. a dozen) windows; these windows are 
of same size and shape; each window consists of an upper U- 
structure and a lower U-structure and each of these consists of 
two L-primitives in according symmetric convex configuration. 
A careful look at the set of primitives given e.g. in Fig. 3 shows 
that only a minority of the windows perceived by humans in 
Fig. 2 allow reduction to instances Rectangle according to this 
system. Most often something is missing or badly displaced. 
Accordingly, the grouping of non-trivial Row and Lattice 
instances will also fail. There is little sense in trying automatic 
interpretation with this system. 
2) Experience shows that often one corner is missing, while 
other corners appear multiply in displaced versions. There is a 
standard approach to cope with such situation: The symmetry 
axes of one vertical U-structure and one horizontal U-structure 
are intersected. Additionally, one side of these structures must 
be quite close to one of the other, and of course again convexity 
is demanded. Thus even incomplete windows can be 
instantiated, and attributed with height and width. But they will 
be instantiated multiply, and for this reason a clustering 
production is included, that fuses several such adjacent Intersect 
instances into one Rectangle object. The rest of the system — 
namely grouping into rows and lattices is the same. All systems 
used here first group in horizontal direction and then in vertical. 
Experiments with this system are reported below. 
3) The third variant attempts to group the window corners into 
rows first. This has the advantage that some of the corners have 
higher probability of appearing than others (according to their 
orientation). Quite long such rows can be grouped, and thus the 
generator vector (shift from one window to the next) can be 
estimated with good precision. Then from two such rows a row 
of U-structures can be built simultaneously with all parts in one, 
and afterwards a row of windows with common width and 
height for all windows which are part of it. So this follows a 
different part-of hierarchy than the one used above. This 
follows the idea that two nearby rows of structures having the 
   
    
      
   
   
  
   
    
    
  
    
     
   
  
    
     
   
  
  
   
same generator with high accuracy probably result from the 
same repetitive pattern. We can be more liberal with biased 
displacements such as shear and un-biased displacements will 
be averaged out by the previous grouping. Experiments with 
this system are also reported below. 
  
Left-hand Right-hand constraint 
U-structure L-primitive, L-primitive symm. & convex 
  
  
  
  
  
  
Rectangle U-structure, U-structure symm. & convex 
Row Rectangle, Rectangle horizontal proximity 
Row Row, Rectangle good continuation 
Lattice Row, Row vertical proximity 
Lattice Lattice, Row good continuation 
  
System "canonical" 
  
  
U-structure L-primitive, L-primitive symm. & convex 
  
  
  
  
  
  
  
Intersect U-structure, U-structure prox. & orthogonal 
Rectangle Intersect, ..., Intersect proximity 
Row Rectangle, Rectangle horizontal proximity 
Row Row, Rectangle good continuation 
Lattice Row, Row vertical proximity 
Lattice Lattice, Row good continuation 
  
2» 
  
System "windows first 
  
  
  
  
   
  
  
   
  
  
  
  
   
  
  
   
  
  
L-Row L-primitive, L-primitive horizontal proximity 
L-Row L-Row, L-primitive good continuation 
U-Row L-Row, L-Row parts(sym. & conv.) & 
similar generator 
Row U-Row, U-Row orthogonal & similar 
generator 
Lattice Row, Row vertical proximity 
Lattice Lattice, Row good continuation 
  
  
  
  
System "L-rows first 
  
  
  
  
Table 1. Production systems 
2.3 Automatic Interpretation 
Search: The grouping uses the interpretation system proposed 
by Michaelsen et al. (2011) which is a successor of the BPI 
system (Stilla & Michaelsen, 1997). Two types of productions 
are feasible: Normal form productions and cluster productions. 
Only one cluster production rule is used here (third of the 
"windows first"), all others are normal forms. Each production 
tests a geometrical constraint on the right hand side objects and 
in case of success infers and assesses a new left hand side 
object. Primitives must be assessed by the extraction process. 
The assessments are important because the search of the 
interpreter is mainly assessment driven. Optional top-down 
acceleration of the search is possible and recommended. The 
search can be terminated either by exhausting all possibilities, 
or after a time limit is reached, or when the first target object is 
found. 
Decision: As result of a search a set of non-primitive instances 
has been accumulated. A decision procedure must be defined 
selecting from these a single or a small sub-set that can serve as 
result e.g. for the next step of the analysis. First one or few 
object classes are picked; here these are Row and Lattice 
objects. From these first the best object is selected; here the 
Lattice instance containing most windows, and among these the 
one that is best assessed by the search process, and if there is no 
lattice than the best Row instance. All instances similar to this 
one are suppressed by local inhibition, and then the next best is 
picked, and so forth. Such rank ordering of accumulated 
interpretation results follows von Hansen et al. (2006). In the