International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004
A very important source often neglected in more theoret-
ical work are GIS data. (Brenner, 2000) use two dimen-
sional (2D) polvgons from which they generate straight
skeletons in conjunction with laser-scanner data to effi-
ciently and reliably extract buildings. In (Gerke et al.,
2004) a two stage process is employed to verify given road
data. After extracting reliable roads in the first stage using
strict parameters, topologic information is used to restrict
the further analysis so that relaxed parameters can be used
leading to a more complete verification and therefore to a
higher efficiency. (Zhang, 2004) employ color and stereo
data together with extensive modeling, comprising, e.g.,
context, occlusions, and shadows, making heavy, yet intel-
ligent use of the GIS data, leading to an impressive perfor-
mance.
Even though the above papers show the potential of using
GIS information, it is essential to keep in mind, that one
cannot rely absolutely on it, as it might be outdated and
unprecise and therefore lead to wrong conclusions. There
is always a trade-off to be made between accepting wrong,
because changed objects, and rejecting many correct, be-
cause unchanged objects. Therefore, even when using ad-
ditional information from GIS, in most cases approaches
are needed which model reality so deeply, that they can ex-
tract the objects at hand even under complex circumstances
as, e.g., (Zhang, 2004) demonstrates it.
2.3 Statistical Modeling
The deficits of a mainly deterministic modeling, for in-
stance based on semantical networks, e.g., (Niemann et
al., 1990), have been known for a long time. There have
been heuristic attempts by adding for instance believe val-
ues, but more sound ways of including statistical modeling
have been used only recently, for instance Bayesian net-
works in (Growe et al., 2000) or (Kim and Nevatia, 2003).
The work on dynamic Bayesian networks (Kulschewski,
1999) has been interesting in terms of modeling objects
and their relations. Though, manually generated ideal data
were used, and thus the feasibility of the approach to cope
with real world, noisy, and unreliable data is hard to judge.
Until recently, semantical modeling was also lacking the
capability to visualize the actual contents of the knowledge
modeled. The quality of the modeling, e.g., by a semanti-
cal network, could only be judged by looking at interpre-
tation results and it was not clear how much a component
contributed to the results.
By the advent of reversible Jump Markov Chain Monte
Carlo (RIMCMC) (Green, 1995) there is a means for sta-
tistical modeling which can also be used for simulation.
The jumps in RIMCMC make it possible not only to use
distributions for the parameters of objects and relations,
but also to introduce new objects or relations and to delete
them. The latter is the reason why the jumps are called
reversible: For every jump generating a new object there
needs to exist a backward jump, allowing to eliminate the
object. Because of this, RIMCMC has the following out-
standing features:
416
e The modeling is extended in a sound way to deal with
the uncertainty of objects as well as their relations
even when it is not known beforehand, which and how
many objects exist.
e It is possible to sample into the distribution allowing
to simulate objects and their relations according to the
model. Thus, one can check from the outcome, if
the given model really describes what it is supposed
to describe. Le., in stark contrast to most modeling
schemes, one can check the model without analyzing
given data.
That the ideas of RIMCMC are practically feasible and
meaningful was shown by work on facade interpretation
(Dick et al., 2002), road extraction (Stoica et al., 2004),
and vegetation extraction (Andersen et al., 2002). The
former two demonstrate that one can produce realistically
looking facades or roads, respectively, by starting from a
few basic primitives, such as a window and a door, or a
road piece, and then sampling into the distribution. For the
roads and the vegetation, sampling is done for the extrac-
tion in conjuction with simulated annealing, avoiding local
minima, but also resulting in a very high complexity of the
approach.
Another issue of statistical modeling is self-diagnosis.
(Forstner, 1996) introduced the “traffic light” paradigm.
Results which are correct (green) are distinguished from
certainly incorrect results (red) and results, which might be
correct, but should be checked (yellow). The idea is that a
calling routine will get back information if it can rely on
a result (green), if the result might be correct (yellow), or
if there was no meaningful result (red). Self-diagnosis is
based on statistical modeling. The more one knows about
the deterministic and stochastic structure of the problem,
the more reliable self-diagnosis will be. (Gerke et al.
2004) have built their approach for road verification on top
of the traffic light paradigm.
2.4 Geometry and Statistics
An area of statistics linked to problems often geometri-
cal in nature is concerned with the large number of blun-
ders in the data, the vision community always has to deal
with, especially when using matching algorithms. This has
sparked the development of techniques which approach the
problem differently from how most photogrammetrists do
this. Especially popular is the random sample consensus,
or short RANSAC approach of (Fischler and Bolles, 1981)
and its variants such as the geometric information crite-
rion (GRIC) (Torr, 1997). The basic idea is to take a larger
number of random samples with the minimum number of
observations necessary to solve the problem. All these
samples lead to solutions which are then checked against
the rest of the observations. Finally, the solution is taken,
which is in correspondence with the largest portion of ob-
servations. This technique is extremely useful for applica-
tions such as the estimation of the epipolar geometry (Hart-
ley and Zisserman, 2000), for aero-triangulation (Schmidt
Intern
and B
3D pc
Comp
proble
well.
while
serma
years
tant ir
by (Fz
hahn :
metric
articul
work
algebr
matioi
25 1
From
autom
els frc
tance
generz
SONS, \
variety
For le:
degree
the ful
as buil
geome
Unfort
been ii
Micha
well w
tion.
large [
ter a lo
unders
Also í
den M
throug
Insteac
(gramr
plicati
cies of
1999).
age pr
progre:
Finally
learnin
discuss
nition :
concep
logical
cipled :
