| guar-
Niev-
oceed-
Verita,
= Vol,
lation
rams.
ch for
s In:
Inter-
il Ge-
roup,
ation
Sci.
Spa-
ams.
fore-
rich-
nna,
tion.
nter-
John
i 2004
ROBUST AND FULLY AUTOMATED IMAGE REGISTRATION USING INVARIANT FEATURES
Joachim Bauer, Horst Bischof", Andreas Klaus, Konrad Karner
VRVis Research Center, Austria
[bauer,karner]@vrvis.at
*Institute for Computer Graphics and Vision, Graz University of Technology, Austria
bischof@icg.tu-graz.ac.at
KEY WORDS: Photogrammetry, Architecture, Geometry, Matching, Feature.
ABSTRACT
This paper introduces a novel method for affine invariant matching using Zwickels that is especially well suited for im-
ages of man-made structures. Zwickels are sections defined by two intersecting line segments, dividing the neighborhood
around the intersection point into two sectors. The information inside the smaller sector is used to compute an affine
invariant representation. We rectify the sector using the line information and compute a histogram of the edge orientations
as a description vector. The descriptor combines the advantage of accurate point localization through line intersection as
well as higher descriptivity through use of a larger image region compared to descriptors computed around the points.
Compared to other affine invariant descriptors we demonstrate that our method avoids the problem of depth discontinu-
ities. In several matching experiments we show that our features are insensitive against viewpoint changes as well as
illumination changes. Results are presented for aerial and terrestrial images as well.
1 INTRODUCTION
The computation of features that are invariant against view-
point and illumination changes is a crucial step in every im-
age matching or image indexing task. Commonly used fea-
tures are the affine invariant ones, since perspective trans-
forms, as they occur in wide baseline setups can be lo-
cally approximated by an affine transform. Typically an
interest point detector provides locations at which a lo-
cal affine invariant descriptor is computed. Based on the
assumption, that the area around the interest point is pla-
nar or sufficiently smooth an affine invariant descriptor is
useful. Several methods have been proposed in literature
e.g. by. Baumberg (Baumberg, 2000), Lowe(Lowe, 1999),
Schmid and Mohr (Schmid and Mohr, 1997). Mikolajczyk
and Schmid (Mikolajczyk and Schmid, June 2003) eval-
uated the performance of several local descriptors. The
most challenging problem in these approaches is to find
the correct scale i.e. the spatial extension of the support
region around the point. Other methods define an invariant
region by finding a stable border as proposed by Schaffal-
itzky and Zisserman ( Schaffalitzky and Zisserman, 2001),
Tuytelaars and Van Gool (Tuytelaars and Gool, 2000) or
Matas et.al (Matas et al., 2002). Larger regions seem to be
preferable because they allow a more distinctive descrip-
tion, but on the other hand are more likely to contain oc-
clusions if the same region is viewed from a different view-
point. Larger regions may also deviate from the planar case
or exhibit large perspective distortion.
In this paper we present a method for the detection and
affine invariant description of image regions using Zwick-
els !. A Zwickel is formed by the intersection of two lines,
where the intersection points of the line segments serve as
interest points. The principal idea behind this approach is,
that the area between intersecting lines is in many cases
planar. Unlike other methods that compute the descrip-
tor for a symmetric or skew-symmetric region around the
! German: zwicken : to nip
1419
interest point, we use the dividing property of the line seg-
ments to compute the descriptor only for the smaller sector.
This has the advantage, that if two sectors match, we com-
pare only the correct parts and thereby achieve a higher
discrimination ability, especially if lines are lying on depth
discontinuities. Our approach is split up into two steps:
first we detect potential Zwickels by searching for inter-
secting line pairs. This step yields accurate points of in-
terest and subdivides the region around this point into two
sectors. The lines therefore automatically provide a seg-
mentation by dividing the region around the interest point
into two sectors.
In the second step we compute affine invariant descriptors
for those sectors that are enclosed by the intersecting lines.
The computation of the affine invariant descriptor involves
a rectification of the enclosed sector and the construction
of a histogram of the edge orientations. It is clear, that
the proposed interest points can only be detected in im-
ages, where a sufficient number of lines is present - this
is true for images containing typical man-made structures.
The geometric accuracy of the intersection points is higher
than those of corner based points of interest. The outline of
the paper is as follows: In section 2 we describe the detec-
tion of Zwickels and the computation of the affine invariant
descriptor. Section 3 shows the application of the Zwickel
descriptors for image matching. Experiments with real and
synthetic images are presented in section 4, concluding re-
marks and an outlook in section 5 close the paper.
2 ZWICKEL DETECTION AND DESCRIPTION
In the following we describe how Zwickels are detected,
explain the rectification process in more detail and address
the computation of the affine invariant descriptor.
2.1 Zwickel detection
The detection of Zwickels is performed as follows: In the
first step 2D line segments are extracted from the image,
