In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009 
233 
3.1 Production of Facade Hypothesis 
The production process starts with an arbitrary facade, called 
the axiom, and proceeds as follows: (1) Select a non-terminal in 
the current string, (2) choose a production rule with this non 
terminal as predecessor, (3) replace the non-terminal with the 
rule’s successor, (4) terminate the production process if all non 
terminals are substituted, otherwise continue with step (1). The 
geometrical result of the production process depends on the 
order in which the non-terminals are selected. Usually, best 
results are obtained when facade structures which are likely to 
appear in the middle of the facade are placed first, and the 
remaining spaces to the left and the right side are filled 
afterwards. As it is illustrated in Figure 7, the non-terminal 
selection refers to this principle. For clearness, we here assume 
a facade with only one floor. In each step, the non-terminal 
selected for the next substitution is marked in red. 
side facing scanners are blue (right scanner) and red (left 
scanner). 
Figure 8. Measured facade points and determined convex 
‘dense area’ (blue rectangle) 
Facade strina 
io: F (polygon) — 
i 
WG W * 
Wg W * 
WG Wg, W *- 
w,GWg, W ^ 
ir g. IV g, w * 
Applied rule types 
F — W 
W— WG W 
G—>g, 
W-* WG W 
W-> H 1 , 
G—>g, 
h-, g w,... g, W 
Figure 7. Non-terminal selection 
As long as the facade string consists of only one symbol, the 
non-terminal selection is trivial. In the third line, substitution 
starts with the non-terminal G in the middle of the string. Ac 
cording to this replacement, the chosen geometry tile g, will be 
placed about in the middle of the facade floor. The following 
replacements are taken from the left to the right of the string. 
When there is only one non-terminal left on the right end of the 
string (see the last line in Figure 7), the left part of the facade 
floor is completely filled with a sequence of wall and geometry 
tiles. At this stage, symmetry can be enforced by substituting 
the remaining non-terminal IT by a mirrored version of the left 
terminal string. If no symmetry is required, the replacement can 
be continued as described before. During the production, non 
terminals are successively rewritten by the application of 
appropriate production rules. When more than one production 
rule is possible for the replacement of the current non-terminal, 
the rule with the highest probability value is chosen. As soon as 
the facade string contains only terminals, the production is 
completed and the string can be transferred into a 3D 
representation. 
As it is visible in Figure 8, the point sampling distance varies 
strongly due to occlusions and oblique scanning views to the 
upper part of the building. For this reason, facades may contain 
areas where no or only little sensor data is available. In such 
regions, an accurate extraction of windows and doors cannot be 
guaranteed anymore. Nevertheless, a grammar based facade 
completion allows for meaningful reconstructions even in those 
areas. The main idea is to derive the facade grammar solely 
from facade parts for which dense sensor data and thus accurate 
window and door reconstructions are available. The detection 
of such ‘dense areas’ is based on a heuristic approach 
evaluating the sampling distances of the points lying on the 
facade surface. In Figure 8 the extracted convex dense area is 
marked by a blue rectangle. Since the inference process is 
restricted to this dense area, a facade grammar of good quality 
can be provided, which is then used to synthesize the remaining 
facade regions during the production step. 
Figure 9. Facade reconstruction for the “Lindenmuseum” 
3.2 Application Scenarios 
Within the production process, the grammar is applied to 
generate hypotheses about possible positions of each geometry 
tile and thereby synthesize facade geometry for given coarse 
building models. This process can for example be used to 
generate facade structure at areas, where sensor data is only 
available at limited quality. Such a scenario is depicted 
exemplarily in Figure 8, which shows a StreetMapper point 
cloud for an exemplary facade acquired during two epochs. The 
colours encode the different scanners mounted on the 
StreetMapper. Points that stem from the upward facing laser 
scanner are marked in yellow; points that are measured by the 
Figure 10. Facade geometry synthesized from grammar library