ındh, J.-O.,
es 201-206,
chstumsver-
i. Master's
an segmen-
odels. /EEE
hine Intelli-
ge Segmen-
26(9):1277-
n. Springer,
ing in belief
Bayesschen
PF, Univer-
mantic mo-
rial images.
itors, Auto-
n Aerial and
Basel.
lustering for
ure models.
Machine In-
EVALUATING MODEL FIDELITY IN AN AERIAL IMAGE ANALYSIS SYSTEM
F. Quint M. Sties
Institute for Photogrammetry and Remote Sensing
University of Karlsruhe
76128 Karlsruhe, Germany
quint@ipf.bau-verm.uni-karlsruhe.de
Commission Ill, Working Group 3
KEY WORDS: Aerial Image Understanding, Model, Knowledge Base, Semantic Networks
ABSTRACT
The purpose of the system MOSES is the automatic recognition of objects in aerial images. To direct the model based
structura! image analysis, one has to evaluate each state of the analysis process. One We present in this article the procedures
used in MOSES to calculate a part of these valuations, the model fidelity, which is a measure for the goodness of match
between the chosen image primitives and the specific model. Metrics defined on a parametric representation of the primitives
are used to evaluate the model fidelity. The results of the image analysis process directed by these valuations are presented.
KURZFASSUNG
Das System MOSES dient der automatischen Erkennung von Objekten in Luftbildern. Zur Steuerung der modellbasierten,
strukturellen Bildanalyse sind Bewertungen des aktuellen Analysezustandes anzugeben. In diesem Artikel werden die in MOSES
verwendete Verfahren zur Berechnung eines Teils dieser Bewertungen, der Modelltreue, vorgestellt. Die Modelltreue ist ein MaB
für die Übereinstimmung zwischen den gewáhlten Bildprimitiven und dem spezifischen Modell. Zu ihrer Berechnung werden
Metriken auf einer parametrischen Darstellung der Primitiven verwendet. Ergebnisse des Bildanalyse unter Verwendung der
vorgestellten Modelltreue werden erläutert.
1 INTRODUCTION
| scene description |
| scene description D EH specific model |
generic model
Understanding of aerial images is one of the most challeng-
ing tasks in computer vision. Due to its complexity, a model
based analysis has been found to be mandatory since sev-
eral years, see e.g. (McKeown et al., 1985), (Nicolin and
Gabler, 1987), (Matsuyama and Hwang, 1990), (Sandakly and
Giraudon, 1994), (Stilla, 1995). In our system MOSES (Map
Oriented SEmantic image underStanding) (Quint and Sties,
1995) we too perform a structural, model based analysis. We
are interested in the recognition of objects in urban environ-
ment using large scale aerial images.
generic model
pum mme mu mm
map domain
scene domain
image domain
generative model
2 MOSES
One of the main characteristics of the system MOSES is that Figure 1: Architecture of the system MOSES
large scale topographical maps are used to automatically re-
fine the models used for image analysis. The architecture of
our system is shown in Fig. 1. The generative model contains 2.1 Map analysis
domain independent, common sense knowledge the system
designer has about the environment. The generic models in
the map domain and in the image domain are specializations
of the generative model and they reflect the particularities of
the representations of our environment in the map and image
respectively. The models contain both declarative knowledge,
which describes the structure of the objects, and procedural
knowledge, which contains the methods used during the map
and image analysis process. As a repository for the models ; ;
semantic networks (Findler, 1979) are used, as implemented The nodes of the semantic network represent objects, parts
by the system ERNEST (Kummert et al., 1993). and subparts of the scene. They are described with attributes,
which in this case mainly contain the geometric properties of
In the first phase, the generic model in the map domain is
used to analyse the map, which is available as a list of di-
gitized contours. The procedure by which map analysis is
performed is similar to the one used in the image analysis
process and will be described in a following section. The
result of the map analysis is a description of the scene, as
far as it can be constructed out of the map data. This scene
description is also stored in a semantic network.
The generative model and the generic models are that part
of the system which is build by the system developer. The
models and scene descriptions described in the sequel are
automatically build in analysis processes. Analysis takes place
in three phases.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
the scene objects. Links between the nodes represent rela-
tions between the corresponding objects or parts. Two typical
relations are the part-of relation, which describes the struc-
ture of the scene objects and the specialization relation, along
which properties of objects are inherited.
669
