(B) using a 
\TIONS 
'roject needs 
. In this sec- 
ited contour 
zht contours 
n photomet- 
ntour graph 
collection of 
s are gener- 
h effort has 
that reliably 
does not re- 
eve it is wise 
> best possi- 
rgy operator 
tor produces 
es in images 
rs due to its 
es and lines, 
ne pixels are 
he algorithm 
Fig. 6B is a 
:cted to each 
-D points. 
on about ge- 
ss, attributes 
he attributes 
the contour 
> denoted re- 
g contour. A 
ur by a trans- 
f its normal. 
nstructed re- 
etails on the 
son 1995]. 
irly homoge- 
data points 
n a small re- 
t also be ac- 
   
  
  
   
   
  
  
  
   
Figure 5: (A) the original (RGB) image of the residential scene, (B) the Digital Surface Model (DSM), (C) the manually 
measured CAD models of buildings, (D) the result of color classification alone. The pixels of the building class are shown (all 
other pixels are black) (E) the result of classification for the MMO class and the projected DSM blobs (in grey), with NO 
shown in black. The upper right house is not included in the DSM. (F) the result of building detection after combining the 
spectral classification and the DSM blobs and refining the outline of the blobs by the use of edges. 
  
Figure 6: (A) a cut-out from 
the original image in Fig. 5A, 
(B) The resulting contour 
graph with all its contours 
and end-points, (C) the 
flanking regions with their 
corresponding mean light- 
ness attributes. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
counted for. The photometric and chromatic attributes are 
computed for each flanking region using the CIE L*a*b* color 
space. The photometric region attributes are computed from 
the lightness component L^ , whereas the chromatic region 
attributes are derived from the a* and b" color components. 
First, the mean lightness and its standard deviation are es- 
timated by applying the Minimum Volume Ellipsoid (MVE) 
estimator [Rousseeuw and Leroy 1987] on the L^ data. The 
inliers in L^ are then used to robustly estimate the mean 
vector and the scatter matrix for the chromatic components 
(a*,b*). Again, the MVE estimator is used, however, with 
two variables (a^, b^). The estimated scatter matrix of the 
chromatic cluster is then diagonalized. The chromatic at- 
tributes are thereby represented by the mean vector and the 
two eigenvalues of the scatter matrix. In Fig. 6C we show the 
mean lightness L* of each flanking region. The photometric 
and chromatic region attributes are used to compute similar- 
ity relations (next section) and in segment stereo matching 
(section 6.1). 
5.3 Contour Similarity Relations 
Although geometric regularity is a major component in 
the recognition of man-made structures, neglecting other 
sources of information that corroborate the relatedness 
among straight contours imposes unnecessary restrictions on 
the approach. A popular means to relate pairs of straight 
325