3b. Beijing 2008 
447 
SELECTING APPROPRIATE FEATURES FOR DETECTING BUILDINGS AND 
BUILDING PARTS 
Martin Drauschke, Wolfgang Fdrstner 
University of Bonn, Department of Photogrammetry 
Nussallee 15, 53115 Bonn, Germany 
martin.drauschke@uni-bonn.de, wf@ipb.uni-bonn.de 
KEY WORDS: Building Detection, Classification,, Feature Selection, Stable Regions, Adaboost 
ABSTRACT: 
The paper addresses the problem of feature selection during classification of image regions within the context of interpreting 
images showing highly structured objects such as buildings. We present a feature selection scheme that is connected with 
the classification framework Adaboost, cf. (Schapire and Singer, 1999). We constricted our weak learners on threshold 
classification on a single feature. Our experiments showed that the classification with Adaboost is based on relatively small 
subsets of features. Thus, we are able to find sets of appropriate features. We present our results on manually annotated and 
automatically segmented regions from facade images of the eTRIMS data base, where our focus were the object classes 
facade, roof, windows and window panes. 
1 INTRODUCTION 
The paper addresses the problem of feature selection during 
classification of image regions within the context of inter 
preting images showing highly structured objects such as 
buildings. Our region classification is meant to be used as 
an initial interpretation which can be inspected by a high- 
level system, cf. (Hartz and Neumann, 2007), leading to 
new hypotheses to be verified by new image evidence. 
The diversity of buildings and their environment is too rich 
for the determination of appropriate region features by only 
a few features, such as the color channels themselves, gra 
dient images or texture measures. This would lead to un 
satisfying results during classification of building parts. If 
we, as a remedy, significantly expand the dimension of the 
feature vector, we need to select appropriate features for 
efficiency reasons. 
This paper adresses the problem of feature selection. We 
present a feature selection method based on Adaboost and 
investigate its performance with respect to classification of 
image regions which show buildings or building parts. In 
the following we first relate our work to previously pro 
posed methods and use it to motivate our approach. Then 
we introduce our technique and decribe our experiments 
with which we want to show the capability of our approach. 
2 RELATED WORKS 
Building extraction is an active research area in photogram 
metry. Concerning building extraction from aerial images, 
the review in (Mayer, 1999) presents well discussed ap 
proaches. Many of these methods work on extracted image 
edges to identify buildings, e. g. (Nevatia et al., 1997), or 
combine detected image edges and detected homogeneous 
image regions in a hierarchical aggregation process, e. g. 
(Fischer et al., 1998). Regrettably, these approches often 
fail as soon as the building structures become too com 
plex. The interpretation of terrestrial images of buildings 
is not developed as far as the interpretation of aerial im 
ages. Some approaches only investigate the grouping of 
repetitive textures and structures, cf. (Wang et al., 2002) 
and (Tuytelaars et al., 2003). 
Building detection is also a very active research area in pho- 
togrammtery and computer vision. In recent approaches, 
graphical models are often used for integrating further in 
formation about the content of the whole scene, cf. (Kumar 
and Hebert, 2003) and (Verbeek and Triggs, 2007). In an 
other paradigm, the bag of words, objects are detected by 
the evaluation of histograms of basic image features from a 
dictionary, cf. (Sivic et al., 2005). Unfortunately, both ap 
proaches have not been tested with high resolution building 
images. Furthermore, the bag of words approaches have 
not applied to multifarious categories as building or mam 
mals. Epshtein and Ullman (2005) and Lifschitz (2005) 
propose a hierarchical interpretation scheme and show first 
results on very small and strongly smoothed building im 
ages. Thus, their concept is also not applicable on detailed 
building scenes. In photogrammetry, the interpretation of 
building facades is currently studied by Reznik and Mayer 
(2007), where they make use of image sequences. 
Feature Selection is used in data mining to extract useful 
and comprehensible information from data, cf. (Liu and 
Motoda, 1998). In our case, we want to select the most 
valuable features from a set of candidates to keep the clas 
sification efficient and reliable. One classical approach, the 
principal component analysis, reduces the dimension of the 
feature space by projecting all features, cf. (Bishop, 2006). 
Thus, the obtained new features are not a subset of all can 
didate features, but combinations of the original features. 
The exhaustive evaluation of all 2 n combinations of D fea 
tures takes too much time. Therefore, we may choose be