following Section 3 the best five instances in this rank order are 
displayed. 
3. EXPERIMENTS 
Images of the temperature on facades in Munich have been 
obtained such as displayed in Fig. 1. All the systems are 
evaluated on the primitives displayed in Fig.2. There is little 
chance in obtaining any usable result with the system 
"canonical" system presented in Tab. 1. E.g. most of the 
windows are incomplete, and some primitives are badly 
displaced. Results of running system 2 and 3 of Table 1 are 
given below. The computational effort has been fixed by 
stopping the search after the same fixed time. The five best 
ranking results are displayed in Figs. 4 and 5. 
3.1 Results with the “windows first” system 
From the 365 L-primitive instances 291 U-structure, 331 
Intersect, 139 Rectangle, 31053 Row, and 0 Lattice objects have 
been inferred. 
Figure 4 displays a result obtained after searching the data using 
the production system that looks for windows by clustering 
intersections of nearby orthogonal U-structures. With our 
preliminary parameter setting this can find almost all parts of 
the salient upper window row (one window in the middle and 
one on the right margin missing). Row gestalts are indicated by 
yellow rectangles connected by a blue line. The best gestalt 
contains nine windows; second best in rank is a four window 
row on the right side, where the generator is found with good 
precision also, but the window sizes are a little too small. Third 
and fourth in rank are coincident with parts of the best, each 
containing only six windows. 
The middle row of windows appears badly disturbed. The fifth 
in rank Row gestalt sees four wide windows there. It guesses a 
generator which is in 4/3 harmony with the correct one. This 
has got to be regarded as failure. Without knowledge on the 
particular form or size of the windows this cannot be avoided. 
Looking on the primitives in Figure 3 only even a human 
observer would be tempted to see such illusory gestalt. Below, 
on the first floor nothing is found. Since neither the generator 
nor the window size matches no Lattice object can be 
instantiated. 
  
Figure 4. Result with “windows first” productions 
3.2 Results with the “L-row first” system 
From the 365 L-primitive instances 23381 L-Row, 9422 U-Row, 
18864 Row, and 176 Lattice objects have been inferred with the 
same computational effort as in Sect. 3.1 — i.e. 300 seconds in 
eight parallel threads and resorting the queue after 64 
hypotheses, using pure bottom-up data-driven control. 
   
   
  
    
    
    
  
    
    
      
    
     
    
    
    
    
     
    
    
    
    
    
   
    
   
    
    
    
     
    
      
   
    
      
  
Figure 5. Result with “rows first” productions 
Figure 5 displays again the best five resulting instances, two 
lattices and three rows. Both lattices contain 2x5 windows and 
the vertical spacing is estimated roughly correct. The upper row 
alone is less complete than in Section 3.1 (one row of eight 
members and one of seven with overlap). But there are also 
results with correct generator and window size on the middle 
row (one row of five members). Still, the phase of the middle 
rows is estimated wrongly; they are all displaced left and a little 
upward. Below, on the first floor again nothing is found. 
4. DISCUSSION AND CONCLUSION 
Obviously, such data are not easy to be parsed. But we can state 
the following: An object with non-trivial part-of structure — 
such as a facade — may be decomposed in different ways using 
different kinds of non-terminal objects in between. Here we 
have given three different decompositions of facade objects 
coded as production systems. It turns out that while the systems 
seem equivalent in the generative right-to-left direction — e.g. 
for use for fagade rendering - they do not yield the same 
behaviour in the reducing direction, i.e. for recognition. In fact 
those systems that code natural, common sense decomposition 
such as “a facade consists of a stack of rows; each row consists 
of windows of equal size; each window consist of an upper and 
a lower U-structure matching; and each such U-structure is 
made up of two matching L-primitives" won't work for 
recognition at all. 
If one is determined to use a production system in reducing 
direction performing recognition by parsing real data — in 
particular data that contain a large portion of additional clutter 
primitives and also lots of omissions, such as from thermal 
mosaics — the decomposition into non-primitives must be 
chosen with care and different possibilities should be tested 
including non-intuitive decompositions such as grouping the 
primitives into rows first and composing the windows 
afterwards simultaneously on all windows of a row. 
The presented results strongly depend on the reliability of the 
corner detector. A huge number of false and missing detections 
of the corners can lead to errors in whole algorithm. Poor 
detection rate of the presented method (Figure 4 and Figure 5) 
is related to the texture reconstruction techniques, which sticks 
many images of a video to one image. In this process small 
distortions cannot be avoided. Considering Figure 3 also à 
human would have difficulties to recognize windows. 
Improving the matching between frames would reduce 
distortions and would deliver better results. 
The results also depend on the constraints given in the 
rightmost column of Table 1. These contain threshold 
parameters to be chosen by an expert familiar with the issue. 
  
The sys! 
where p 
the othe 
lead to 
paramet 
represer 
Minor f 
paramet 
parallel 
latter is 
And the 
little ini 
setting 
influenc 
41 Ou 
More € 
(and la 
principl 
original 
displac: 
Missing 
prolong 
about tl 
Then a 
values 
values | 
The a 
should 
accepta 
row. À 
improv 
Refere 
Ali, H. 
Windo 
Image. 
Ali, H 
detectic 
Signal 
Burock 
unsupe 
image 
Extraci 
Monitc 
Foerstt 
(2009) 
Made | 
von H 
Cluste: 
buildir 
L, pp. 
Hoegn 
and by 
casting 
S (ed 
Intern: 
Spatial