many
rammetric
important
asis. Here,
10w useful
ine. As the
mensional
some peo-
FT opera-
1 Invariant
affine dis-
lefeys and
shown that
t the pose
h the only
so demon-
ing, an is-
)3) for ro-
vital aerial
r the five-
inally, the
) matching
for match-
ique (Kol-
arity esti-
ias, 2004)
actical use
> contents
, yet giv-
paper self-
1] imagery
the former
and other
video se-
pe, we do
| is model-
's also the
'raphic in-
ithout ge-
itific work
valuate the
! compari-
hat testing
object ex-
hich mod-
so shows
a SUCCESS:
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
———Bá
ful test, the next step is to design the user interaction in
semi-automated systems. We discuss the state of the art
and several issues in Section 4. As all technical develop-
ments are nothing without markets, we give in Section 5
an idea about future markets and what other areas. partic-
ularly visualization from computer science, envisage. The
paper ends up with conclusions.
2 MODELING
Modeling is the key issue for the performance of any ap-
proach for automated or also semi-automated object ex-
traction. Basically, modeling consists of knowledge about
the objects to be extracted. Additionally, in most cases it
is necessary to analyze their mutual spatial and topologic
relations as well as their relations to additional objects,
which a customer might not be interested in to extract, but
which give important clues for the recognition of an object.
E.g., even though one is just interested into roads in city
centers, one will only find them, when one knows, where
the cars are (Hinz, 2003).
The modeling of the objects is the key issue. But instead
of analyzing the assets and drawbacks of individual ap-
proaches, as, e.g., in (Mayer et al., 1998, Mayer, 1999), we -
will in the remainder of this paper concentrate on a num-
ber of issues we consider as important to improve object
extraction for DPW. Overall it is our firm believe, that only
by a detailed modeling of many objects and their relations
of the scene, it will be ultimately possible to mostly reli-
ably extract objects from imagery, laser-scanner data, etc.
2.1 Strategy and Multiple Scales
Even though the objects and their relations are the neces-
sary core of modeling, experience shows, that the sequence
of operations employing the knowledge about the objects
and their relations is a, often even the key factor for an ef-
ficient, but also powerful extraction. E.g., it is well known
that markings are an important clue to find roads. Un-
fortunately, in 1mages with a ground pixel size of about
0.25 m the markings very often correspond to very faint
bright lines. When trying to extract them in open rural
space one will in most cases extract millions in the fields
and meadows leading to an infeasible grouping problem.
On the other hand, one can first produce hypotheses for
roads in the form of lines in images of a reduced resolu-
tion, i.c., images in a higher level of an image pyramid.
Then one verifies the roads in the form of directed homo-
geneity such as in (Baumgartner et al.. 1999). Inside the
tracted and grouped reliably, giving the hypotheses a high
evidence for being actually roads.
We term the basic concepts behind a sequence of opera-
tions controling the extraction the "strategy". Ideally, there
exist objects
e which are easy to extract,
e can be extracted reliably, and
generated hypotheses for roads the markings can be ex- -
a
e which have a large positive influence on the interpre-
tation of the whole scene.
The idea is to find cues for objects which allow to focus
the attention to specific areas, such as hypotheses for roads
to extract markings (cf. above). Unfortunately, this kind of
objects does not always exist and if so, they are not always
easy to identify.
In the above example on roads, scale plays an important
role. Coarse to fine approaches have long been used in
orientation determination and in image matching (Heipke,
1995). For linear objects it was shown in (Mayer and Ste-
ger, 1998), that by means of changing scale from fine to
coarse by means of linear scale-space (Lindeberg, 1994),
one can often eliminate interfering objects such as cars and
trees together with their shadows from roads. Other means
are irregular pyramids, as, e.g., implemented in eCognition
of Definiens GmbH (Benz and Schreier, 2001). A com-
parison of different means is given in (Blaschke and Hay,
2001).
Our experience is, that a multi-scale approach is in many
cases useful. Depending on the type of object, smoothing
with the linear scale-space, eliminating interfering details
by means of gray-scale morphology (Kóthe, 1996), or a
combination of both such as 1n (Kimia et al., 1995) is most
suitable.
2.2 Data Sources and GIS Data
DPW have included in recent years means to deal with high
resolution satellite imagery such as IKONOS or Quick-
bird together with aerial imagery, possibly digital, e:g.,
from Leica‘s ADS 40 (Fricker, 2001), Vexcel‘s Ultracam
(Leberl et al., 2003), or Z/I imaging‘s DMC (Hinz et al.,
2001).
To use data which comprise explicit information suitable
for the problem can be a very efficient means to make
extraction more robust and reliable. These are most im-
portantly color, or more generally spectral data, as well
as three dimensional (3D) data. (McKeown et al., 1999,
Mikhail, 2000) show the advantages of using aerial hyper-
spectral data allowing for reasoning about the materials of
the objects. Both make also use of DSM.
For 3D, highly reliable data from laser-scanners are the
data source of choice. Early experiments with the extrac-
tion of buildings from laser-scanner data where done by
(Weidner and Forstner, 1995). Recently, laser-scanner data
are more and more fused with aerial imagery. For it, the
establishment of a common reference frame plays an im-
portant role to arrive at rich features (Schenk and Csatho,
2002). Work such as (Rottensteiner, 2003) uses addition-
ally to the integration with aerial imagery sophisticated
segmentation methods and a consistent model estimation.
In (Straub, 2003) DSM data from laser-scanners partially
together with reflection properties in the infrared are used
for the extraction of individual trees.
