anbul 2004
| over the
they work
or testing
r number
eal world
aration of
> than the
cept : sta-
nce of the
e strategy
ictive sys-
d level of
, it is not
will have
anges for
chains of
positively,
ontrol in-
are useful
object ex-
computer
s case for
oads, this
Li, 1994,
roads are,
is a good
, in urban
s, they are
non-real-
at the hu-
any cases
-line pro-
any case
of work-
1iportantly
cedure by
he overall
stem use-
The basis
ig. For an
ate trade-
ability. It
jn time to
d idea, to
where the
mplicated
| is not the
additional
of human
extremely
ually cor-
iot have to
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
be checked any more. For the “yellow” results, the situa-
tion is more complicated. It should be avoided to offer the
operator a lot of objects, which are plainly wrong. Also
results with complex topologic errors, which might take
more time to heal than to acquire the whole feature manu-
ally, should not be presented to the operator. Helpful might
be, though, to offer a number of choices, one of which is
with relatively high probability correct.
An efficient semi-automated system should comprise real-
time tools, which help to improve the results obtained fully
automatically. The best way, yet needing also the most ef-
fort to implement and again depending on the current state
of automated systems, is to make use of the results of au-
tomated extraction.
Finally, testing, this time on a very practical level, comes
into play again. Only by customizing the system for spe-
cific customers working on real data will make clear the
strengths but also weaknesses of the whole complex chain
of (semi-) automated object extraction approaches. The
latter includes models and strategies as well as their in-
tegration with suitable work-flows and real-time tools for
user interaction. The overall goals are maximal efficiency
and, often even more important, minimal cost.
Because of the large costs, the high risks, the dependence
on in-depth knowledge, as well as on specific production
environments to be tuned for, practical semi-automated ob-
ject extraction is and will be in many cases first developed
in cooperation of academia and data producers, especially
NMA. Only after reasonable success and especially versa-
tility will have been demonstrated, the main DPW devel-
opers will probably join in. Yet, even the above coopera-
tion of academia and NMA on a larger scale would be a
large achievement, because as (Baltsavias, 2004) notes, at
academia there is often a “lack of practical spirit.”
5 APPLICATION AREAS
The traditional market for DPW consists of the acquisi-
tion of 3D topographic information, such as buildings and
roads. DPW have included means to efficiently handle
high resolution satellite imagery together with aerial im-
agery, but also multi- and hyperspectral as well as laser-
scanner data. Yet, to our knowledge there is not a strong
tendency to integrate also tools to handle terrestrial im-
agery or close-range laser-scanner data. We think that this
is a deficit and we will explain why as well as which addi-
tional application areas we see especially concerning veg-
etation in the remainder of this section.
That close range data is not considered in DPW is in con-
trast to a recent issue of the IEEE Journal of Computer
Graphics and Applications focusing on 3D reconstruction
and visualization (Ribarsky and Rushmeier, 2003). The
paper starts with the statement “We have entered an era
where the acquisition of 3D data is ubiquitous, continuous,
and massive" Highly detailed 3D city models from high
resolution terrestrial images, dense video sequences, and
419
terrestrial laser-scanner data are seen to be useful for vir-
tual television, tourism, but also mission rehearsal for fire
fighting or security and rescue scenarios.
Even though there is one photogrammetric paper (Rotten-
steiner, 2003) on building extraction from laser-scanner
data also in conjunction with aerial imagery in the above
[EEE journal issue, the survey on large-scale urban model-
ing (Hu et al., 2003) shows, that the awareness of the work
done in photogrammetry is not too big. As usual, this can
be only changed by submitting papers in this area, but also
by going to the particular conferences.
Recently, there is a large interest into producing highly de-
tailed 3D city models. One of the first and largest projects
in this area is the city-scanning project at MIT (Teller,
1999). Two of the most advanced approaches using im-
ages only are (Dick et al., 2002) and (Werner and Zisser-
man, 2002). (Dick et al., 2002) use advanced statistical
modeling in the form of RIMCMC (cf. Section 2.3) allow-
ing for the reconstruction of complete models from partial
samples of the object. (Werner and Zisserman, 2002) show
what can be achieved assuming that an object is made up
of planes (facades or roofs), which are partially vertically
oriented, have some parallel structures in front of them
(columns) or behind them (windows, doors), and which
can be symmetrical (roofs of dormer window).
Terrestrial laser-scanners can be used for very complex his-
toric sites (Allen et al., 2003). Other approaches combine
terrestrial and aerial imagery as well as laser-scanner data
to produce 3D models with a good fidelity seen from the
top but also from the ground (Früh and Zakhor, 2003).
An area where not too much research has been done is the
extraction of vegetation outside forests, especially in cities.
While it is useful information for city administrations, it is
extremely important for the generation of realistic visual-
izations. (Bacher and Mayer, 2000) use the shadow pro-
jection of the tree in an aerial image together with the fact,
that the vertical trunk of the tree points to the nadir point.
In (Andersen et al., 2002) RIMCMC is used to find trees in
aerial laser-scanner DSM employing knowledge about the
sensing process as well as the spatial interaction of indi-
vidual trees modeled by a Markov process. (Straub, 2003)
model the shape of trees to extract them from aerial laser-
scanner DSM possibly together with reflection properties
in the infrared.
Even less work on vegetation extraction has been done in
the close range. In (Shlyakhter et al., 2001) the hull of
the tree is determined from its silhouette in several im-
ages. The 3D medial axis is computed for the hull and
from it a representation based on an L-system (Méch and
Prusinkiewicz, 1996) is derived. By this means one is able
to derive a tree model with which one cannot only deal
with occlusions, but which can also be animated, e.g., to
simulate wind, and which can be adapted to the seasons.
And, the model corresponds closely to the actual tree at
the given position.
