The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
facade 
5 10 15 20 25 30 
index of iteration 
5 10 15 20 25 30 
index of iteration 
sky vegetation 
Figure 4: Development of the number of used features c\ 
(dashed blue line) and the maximum frequency C2 (solid 
red line) with respect to the 30 weak classifiers h t . 
on features. For window classification, the weak classifier 
mainly use the three features /5 (in 28 cases!), /4 and /is- 
Although the classification of window panes yields much 
better results, the curves of Ci and C2 change very fast, cf. 
fig. 5. Nevertheless, the histograms show that only a small 
subset of four features is mainly used by the 30 weak clas 
sifiers, cf. fig. 6. 
window window pane 
Figure 5: Development of the number of used features ci 
(dashed blue line) and the maximum frequency C2 (solid 
red line) with respect to the 30 weak classifiers ht. 
5.2 Experiments on segmented data 
We determined the stable regions from 82 facade images 
from Berlin and Munich, Germany, and Prague, Czech Re 
public. We changed the image data that was the basis of the 
previous experiments a bit, so we only worked with build 
ings of similar scale and with images of these buildings that 
have a similar resolution. We extracted over 13 000 stable 
regions, and then we repeated the cross validation tests on 
five selections of regions. First, we considered all scales, 
and then we only considered regions which where stable in 
the scale-space layers with scales a = 1,2, 4, 8, respec 
tively. 
Figure 6: Ffistograms over the features of each weak clas 
sifier h t \ /2, /5 and /13 are the most dominant features for 
classifying window panes. 
regions of the same scale. When choosing regions from all 
scales, all weak classifiers choose one of the three features 
/1, /2 and /5. Thus, ci < 3Vf, cf. fig. 7. Surprisingly, the 
bad classification results of roof regions in scale er = 1 is 
also based on stable decisions of the weak classifiers since 
only 11 features are used by all weak classifiers, cf. fig. 8. 
Table 5: Error rates on automatically segmented regions. 
scale 
facade 
roof 
sky 
vegetation 
all 
39.67% 
15.88% 
44.44% 
20.75% 
1 
35.00% 
81.06% 
44.81% 
33.17% 
2 
42.94% 
28.44% 
47.55% 
45.27% 
4 
53.96% 
39.76% 
50.43% 
43.26% 
8 
44.71% 
25.14% 
51.71% 
58.86% 
all scales scale <7 = 1 
Figure 7: Development of the number of used features c\ 
(dashed blue line) and the maximum frequency C2 (solid 
red line) with respect to the 30 weak classifiers h t . 
6 CONCLUSION AND OUTLOOK 
In this paper, we presented a feature selection scheme which 
is connected to the classification framework of Adaboost. 
We chose very simple weak classifiers which only work on 
single features. Thus, we were able to derive information 
on the relevance of features for the classification process. 
We could show, that the weak classifiers favour the use of 
the same features. So, we obtained sets of appropriate fea 
tures even for the cases where we did not find good classi 
fication results. Therefore, we resume that we should im 
prove the classification, e. g. by using more complex weak 
classifiers on 2-dimensional feature planes as fd x /J. Ad 
ditionally, we should expand the feature space, further fea 
tures might by derived from texture. 
In the tabs. 5 and 6, we show the results of our classifica 
tions. With respect to the experiment on classifying roofs, 
we are surprised that the classification of regions over all 
scales yields much better results than the classification of 
ACKNOWLEDGEMENTS 
This work was done within the project eTraining for In 
terpreting Images of Man-Made Scenes (eTRIMS), STREP