The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
we perform Adaboost with only 30 iterations. We checked, 
that the classification error on the training samples is more 
or less constant after a few iterations and decreases very 
very slowly in further iterations. 
We are not only interested in reasonable classification re 
sults, but we also want to derive the most appropriate fea 
tures. Therefore, we analyze the stability of the decisions 
of the weak classifiers. We repeated our tests V times, 
and then we checked the features the weak classifiers h t 
use. Therefore, we determined histograms over the features 
which are used by the f-th weak classifier, t = 1..30. Fur 
thermore, we focused on two certain characteristics. First, 
we analyzed, how many different features are used by t-th 
weak classifier in V trials: ci, 1 < c\ < D. And secondly, 
we evaluated, how often the best feature has been used by 
f-th weak classifier in V trials: C2, 1 < C2 < V. Both 
characteristics can be directly derived from the histograms: 
ci is the number of histogram entries which are not 0, and 
maximum frequency of a feature, C2, is the value of the 
highest peak. 
If the number of used features is relatively small, then the 
weak classifiers always select the same features. We have 
found a set of appropriate features, if it stays consequently 
below a certain threshold. 
5.1 Experiments on annotated regions 
We selected 62 facade images from Bonn, Germany, and 
Prague, Czech Republic. Together, there are 5284 anno 
tated objects in these images. Over 70% of them are win 
dow panes, and over 20% of the annotated objects are win 
dows. 
5.1.1 Facade, Roof, Sky and Vegetation The first ex 
periment included 35 annotated facades, 37 roofs, 47 times 
sky and 70 times vegetation, i. e. 189 samples in total. 
Since the data set is quite small, we were able to perform 
four Leave-one-out-tests, i. e. V = 189, with the samples 
of one class as foreground and the samples from the other 
three classes as background. The classification errors are 
shown in tab. 2. 
Table 2: Error rates of manually annotated regions. 
facade 
roof 
sky 
vegetation 
9.5% 
48.1% 
1.1% 
11.1% 
In tab. 3 and fig. 3, we present the histograms over the 
features which are used by the f-th weak classifier regard 
ing the classification of facades and roofs. With respect to 
the classification of facades, we notice that the first weak 
classifiers hi to /14 distinguish facades from background 
by using nearly the same features over all 189 trials. Thus, 
the set of appropriate features should certainly contain the 
features /3, /4, / 6 and /163. Then, number of used fea 
tures, ci, increases for the further weak classifiers. Finally, 
in the 30-th iteration ci reaches a value of 21, i. e. the weak 
classifiers h3o of all 189 trials use 21 different features. In 
91 cases, the /3 was chosen, and /1 was chosen in 35 cases, 
again. All other features which are used by /130 only play a 
minor role. 
Table 3: Used features and maximum frequency of a fea 
ture with respect to the weak classifiers h t for classifying 
facades in 189 trials. 
hi 
hz 
hz 
hzo 
h : 187 
/3:2 
/4 : 188 
/i:l 
/144 : 174 
/141 : 8 
/a:91 
/1 :35 
Ci = 2 
c 2 = 187 
ci = 2 
c 2 = 188 
Ci = 6 
C2 = 174 
ci = 21 
c 2 = 91 
The almost perfect classification of sky relies on the fea 
tures /9, /151 and /130. Here, even the last weak classifier 
hzo choses in 164 or 87% of all cases the feature /129, all 
other features which are used by fi.30 can get neglected. Re 
garding the classification of vegetation, the first weak clas 
sifiers mainly use the features /9, /5 and /19. The devel 
opment of the histograms of the further weak classifiers is 
similar to the observation which we made after the facade 
classification. 
The classification results of roofs is inacceptable, and here 
the selection of appropriate feature is not so stable as in the 
other three tests. Although the first weak classifiers always 
uses /15 as the discriminative feature, all further weak clas 
sifiers vary much more with respect to the selected feature. 
We show the histograms of four weak classifiers in fig. 3. 
ci = 1 ci = 4 Ci — 6 ci = 19 
c 2 = 189 c 2 = 140 c 2 = 123 c 2 = 47 
Figure 3: Histograms over the features of each weak clas 
sifier h t and the number of used features ci and the maxi 
mum frequency C2 from classifying roofs. 
Due to the limited space, we do not present additional his 
tograms in detail. But the plots of fig. 4 shows how the the 
number of used features ci and the maximum frequency C2 
develop with respect to the classification target. Comparing 
the curves of C2 with respect to roof and sky, it decreases 
much faster with bad classification results. 
5.1.2 Window and Window Panes In the other two ex 
periments, we tested the classification of windows and win 
dow panes. Since we could use 5284 annotated regions, we 
performed a cross validation test, where we used 10% of 
the regions for testing and the rest for training the classi 
fiers. Then, we repeated all tests V = 20 times. The results 
are presented in tab. 4. 
Table 4: Error rates of manually annotated regions. 
class 
min 
median 
max 
total 
window 
15.4% 
24.7% 
75.3% 
41.8% 
window pane 
3.1% 
6.0% 
70.0% 
10.1% 
The classification results of the windows and window panes 
make us very optimistic that we might get reasonable sets 
of appropriate features. Perturbingly, the classification er 
rors vary too much between the 20 different tests. Further 
more, we cannot find any correlation between the classifica 
tion results and the stability of the weak classifierss choice