The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B3b. Beijing 2008
Figure 1: Rectified facade image and an detail from
eTRIMS data base with manually annotated regions, e. g.
window panes, vegetation.
We did our first experiments on manually annotated regions
from rectified facade images, see fig. 1. These tests can
show us the relevance of the features with respect to an op
timal image segmentation.
In the second experiments, we used automatically segmented
image regions. We obtain these image regions from the
analysis of the image’s scale-space with S discrete layers,
which we have already used in (Drauschke et al., 2006).
We adopted some parameters, i. e. we only consider the
original image and additional 41 layers in scale-space with
scales between a = 0.5 and a = 8. Then, we automati
cally trace the regions through the scale-space structure to
derive a region’s hierarchy, similar to the approach of Bang-
ham et al. (1999). For complexity reason, we reduce the
number of regions by selecting only stable regions in scale-
space structure. Distinct region borders are often good ev
idences for stable regions. Thus, most stable regions cor
respond to man-made structures or are caused by shadows.
The process of determining stable regions is explicitly de
scribed in (Drauschke, 2008). Fig. 2 shows all detected
stable regions in scales a = 2.
Figure 2: Segmented stable regions at scales a — 2.
vectorized region’s border, which is a simplification of the
region’s boundary by using the algorithm of Douglas and
Peucker (1973).
Table 1: List of derived features from image regions.
/1
area
/2
circumference
/3
form factor
/4
vertical elongation of bounding box
/5
horizontal elongation of bounding box
fe
ratio /1 : (/4 • /5)
/7-/12
mean color value in original image
regarding the six channels
/13-/18
variance of color values in original image
regarding the six channels
/19-/108
normalized histogram entries of gradients
magnitude, 15 bins per channel
/109-/156
normalized histogram entries of gradients
orientation, 8 bins per channel
/157
portion of lengths of parallel edges
in vectorized region’s boundary
/l58
portion of number of parallel edges
/l59
portion of lengths of boundary edges
which are parallel to region’s major axis
/l60
portion of number of boundary edges
which are parallel to region’s major axis
/l61
portion of lengths of boundary edges
which are parallel to region’s minor axis
/l62
portion of number of boundary edges
which are parallel to region’s minor axis
/l63
portion of number of orthogonal angles
in vectorized region’s boundary
/l64
portion of lengths of boundary edges
which are adjacent to orthogonal angles
The targets y n are obtained differently. For manually anno
tated regions, we additionally select the appropriate class.
Otherwise, the automatically segmented regions inherit the
class target from the best fitting manually annotated region.
The best fitting annotated region Ai* is determined by
RnAi
* =argmax -flirs? (2)
where R is a automatically segmented region and Ai are all
manually annotated regions. Furthermore, the best fitting
annotated region must fulfill the condition
RV\Ai*
R
> 0.5,
(3)
otherwise the class target of the segmented region will be
set to none, and the segmented region will be always treated
as background.
5 EXPERIMENTS
We derive 164 features from each manually annotated and
each automatically segmented image region. Thus, our sam
ples x n are 164-dimensional feature vectors. These fea
tures are roughly described in tab. 1. Color features are
determined with respect to two color channels RGB and
HSV. Last, we derive the features /157 to /i64 from the
The goal of our experiments was to find appropriate fea
tures from the set of image features, which can be used for
classifying our automatically segmented stable image re
gions. Therefore, we designed the Adaboost algorithm as
follows. First, we only use weak classifiers which perform
threshold classifications on a single feature. And secondly,
