Stephan Scholze 
  
Let us now define the pointwise chromatic similarity measure h. The pointwise chromatic similarity measure h is the sum 
of the normalized cross-correlation in the individual bands a* and 6*, evaluated at the point p — (z, y) in the first view 
and its corresponding point p/ — (z, y^) in the second view 
h = ncc(p, p^) 4- nce(p. p^): (7) 
with ncc;(p, p^) being the normalized cross-correlation 
EN 3 aew [EG dy 4-3) — k(z,y)] [KK (x +i,y +5) — &'(2',y')] 
: à T 2 ; = = 
VEGaew [k(@ +4, y +14) — k(z,y)] V Den [A (z' 4- à, y! - i) — k' (a y')] 
of the intensities in color band k € {a*,b*} for a given correlation window W € {W;, W,.}, as depicted in Figure (4, 
b). The pointwise similarity measure 7; is obtained by correlating the left and right sides of the line segments separately, 
taking into account only the side(s) with satisfy the statistical chromatic similarity constraints. More precisely, the pixels 
actually included in the correlation-mask are selected the same way as described in Section (3.2.1): the included pixels 
only approach neighboring contours up to a specified offset and L*-outlier pixels are not considered. 
  
ncci(p, p^) (8) 
  
For the further reduction of match candidates, we employ the average h of the pointwise similarity measure h along the 
common part of both line segments. A pair is kept, if ^ is positive on the side(s) which satisfied the statistical chromatic 
similarity constraints. 
Until now, we restricted ourselves to use only two input images. The benefits of using three ore more views can be seen in 
the results presented in (Moons et al., 1998, Baillard et al., 1999). One advantage in using three views is the availability 
of a strong geometric constraint: the trifocal tensor enables the prediction of a line segment reconstructed from two views 
in the third view. In this paper, we don’t exploit the trifocal constraint; the third view only enters in the computation of the 
cross-correlation measure. If the third view is selected from a parallel flight line, as sketched in Figure (5), we can already 
overcome the limitations in delimiting the image areas used for cross-correlation. As mentioned above, the orientation of 
the sides €; ;/, and c» ;/, Of the image area A; /, in Figure (4, b) is given by the epipolar lines through the common part 
b of 1. For the cases, where the angle between the line segment in image one and the epipolar line from image two is less 
than 40 degrees, we match this line segment against the line segments from image three. 
Image 3 
Image 1 * Image 2 
Figure 5: Schematic representation of the given input image configuration when considering a third view. The orientation 
of the epipolar lines is sketched with dashed lines. 
4 RESULTS AND CONCLUSION 
We tested our algorithm on a state-of-the-art-dataset, produced by Eurosense Belfotop n.v. It consists of high resolution 
color images of densely built up urban areas in Brussels. The image characteristics are: 1:4000 image scale and geo- 
metrically accurate film scanning with 20 microns pixel size (corresponding 8 x 8 cm? on ground) , four-way image 
overlap, and precise sensor orientation. For visualization purposes the results obtained by the succeeding pruning steps 
are depicted in Figure (6) only for one building. In case (6, c, bottom row), the matched lines stemming from image two 
and image three were merged together in the right hand side picture. 
The algorithm was applied to cut-outs with a size of approximately 500 x 500 pixels. The number of extracted line 
segments with a minimum length of 15 pixels, using the Maximum method described above was about 350 in each view. 
If we only restrict the maximal angle difference and the maximal allowed disparity to reasonable values, the number of 
all geometrically possible pairs is about 3700. This number could be reduced to nearly 1000, using only the the integrated 
chromatic properties of the flanking regions. If we additionally apply the chromatic cross-correlation constraint, we 
achieve a further reduction to approximately 700 candidates, also without loosing a single correct candidate. 
The results demonstrate the efficient and reliable complexity reduction achieved by exploiting color information for line 
segment matching. Future work will focus on integrating a third view — and if available further views — seamlessly 
into the presented framework. We expect an even stronger discriminative power of the method by bringing together 
trifocal constraints and the developed chromatic similarity measures. The matched line segments, in conjunction with the 
extracted chromatic neighboring information will form the starting point for the following reconstruction steps towards 
the automatic generation of 3D city models. 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 821 
EE Ce 
ES ET SERIES ep A 
i 
; 
|