TT
the object
ted DTM
CS Which
)Ving the
ti-sensor.
POT data
anning in
0togram-
g Extrac-
irkhäuser
3D. Com-
d Spiller,
of Photo-
al. (eds.):
7
ional Ar-
Norkshop
Network
| Images.
N orkshop
om à case
de (NL),
ammetric
ver, CD-
base revi-
Gruen, A
tion mod-
Stephan Scholze
EXPLOITING COLOR FOR EDGE EXTRACTION AND LINE SEGMENT STEREO MATCHING
IN HIGH-RESOLUTION AERIAL IMAGERY
Stephan SCHOLZE, Theo MOONS, Frank ADE, Luc VAN GOOL
Swiss Federal Institute of Technology
Communication Technology Lab
Computer Vision Group
{scholze, ade, vangool } @vision.ee.ethz.ch
Theo.Moons @esat.kuleuven.ac.be
Working Group III/3
KEY WORDS: Image matching, Multi-spectral data, Automation, Edge extraction, Buildings.
ABSTRACT
This paper investigates into the added value of color information for edge extraction and straight edge segment matching
between stereo views. For edge extraction in color images different methods proposed in the literature are evaluated
and compared, paying special attention to significance and completeness of the obtained edge-map. To find related edge
segment pairs in different views, we apply an odd-man-out scheme: starting with all geometrically possible pairs we first
rule out pairs, for which the chromatic information provided by the regions flanking the edge segments is incompatible.
To further restrict the number of pairs we compute a chromatic similarity measure based on cross-correlation in the color
bands. Both steps result in a significant reduction of candidate pairs, yet no correct pairs get lost. A main application of
our technique is for automatic 3D building reconstruction from high resolution aerial images.
1 INTRODUCTION
The automatic reconstruction of 3D city models from high resolution aerial images is still an area of active research. The
currently available color images of dense urban scenes have nowadays a resolution of 10 x 10 cm? on the ground or
better. This huge amount of data with its immanent information is on one hand the key to better solve the correspondence
problem, on the other hand it holds enormous complexity. This complexity might be the reason, why color information
was not used to its full extent until now.
Our strategy to sytematically reduce the complexity of the problem is based on an iterative scheme. After each reduction
step a more sophisticated elimination procedure can be applied, since the number of remaining combinations gets smaller
after each step. Since the vast majority of man-made objects, especially building roofs, are of polygonal shape with
straight lines delimiting the roof edges, we find it natural to base our method on straight line segments. Thus our method
could be labeled feature-based. Moreover, it is fully based on color information; at no time we refer to a grey-level image
às data source. The main contributions of our work are: edge detection in the color images and the use of chromatic
information for the stereo matching algorithm.
For edge extraction in color images we have evaluated and compared two different methods proposed in the literature.
The key problem is the fusion of the gradient information from the individual RGB-bands. The maximum of the norms of
the gradients in the three bands turns out to be a good compromise between computational simplicity and completeness
of the edge map (Section 2).
For 3D building reconstruction straight line segments at roof edges need to be matched between overlapping views. This
task is not straightforward. First, the edge detection step will always produce incorrect results to some degree. True
Image features might be detected only partially, continuous contours in the image might get fragmented, even cases where
Important image features are not detected at all can happen in real world applications. Second, due to the weak epipolar
constraint, we will find a huge number of possible pairings between extracted edges in different views. A calculation of
all geometrically possible 3D line reconstructions would therefore yield mostly futile results. To overcome the geometric
ambiguities, we take into account the color of regions flanking the extracted line segments. By comparing the integrated
Color distribution in the flanking regions of putative pairs, we determine a statistical chromatic similarity measure. This
Step reduces the possible combinations to typically 30 percent, before calculating the 3D reconstruction. For the remaining
Pairs we calculate the 3D information. A further significant reduction to about 20 percent is now achieved by computing
Across-correlation based chromatic similarity measure, exploiting the pixel-wise one-to-one correspondence induced by
'Pipolar geometry (Section 3).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 815