Matching of relational descriptions leads to search trees
(Vosselman, 1995). Finding the optimum matching result
corresponds to finding an optimal path through that
search tree, assuming that a cost function is associated
to each leaf of that search tree. The trees involved in
image matching can become very extensive even for two
images. Things even get worse when more images are
used. This is the reason why the number of possible
correspondencies has to be reduced considerably. Still,
no complete search tree using topology in more than two
images shall be generated due to the high computational
cost of that method (Vosselman, 1995).
In a first step, hypotheses of correspondence between
features of all possible image pairs are generated. After
that, the hypotheses of all image pairs have to be
combined in order to consider all images at one time.
(Tsingas, 1992) gives a graph - theoretical approach for
the detection of hypotheses in more than two images
which makes use of heuristic search tree methods. This
approach is designed for aerial triangulation, where no
orientation parameters are available. The method might
. become easier and faster if these parameters are
assumed to be known, as it happens with our application.
A correspondence hypothesis between features in image
space leads to a hypothesis for a surface point in object
space. Many matching algorithms evaluate correspon-
dence hypotheses by assuming the object surface to be
smooth and eliminate hypotheses which contradict to that
model by a robust estimation technique , e.g. (Krzystek,
1995). We will also assume the object surface to be
smooth in a first step. If the image data do not fit that
model, another surface model should be assumed. This
means that a knowledge base of different object models
which can be formulated in object space has to be
developed. However, the assumption of another object
model might allow different possibilities of correspon-
dence between features of different images so that the
generation step might have to be repeated (indicated by
the broken line in figure 3). Again, we want to use
topology to generate more complex object models. Up to
now, first tests regarding the formulation of rather simple
object models have been made using the program
system ORIENT. However, this point remains the
probably most important one for research in our concept.
4. CONCLUSION
Our concept for 3D object reconstruction aims at
developing a feature based matching using topology.
More complex object models should be formulated in 3D
object space, which might give us the possibility to
overcome the problem of occlusions. The geometrical
constraints necessary to reduce the computational
complexity of matching are provided by the integration of
the bundle block adjustment system ORIENT. Basic.
modules, e.g. the data interface to ORIENT and the
creation and handling of image pyramids have already
been implemented in C++. Feature extraction is in the
implementation phase, and first tests regarding the
formulation of object models have been run. The whole
development is closely connected to the new SCOP
environment (Molnar et al., 1996).
696
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
5. REFERENCES
Ackermann, F., 1994: High Precision Digital Image
Correlation. In: 39" Photogrammetric Week.
Schriftenreihe des Instituts für Photogrammetrie, Heft 9,
Universität Stuttgart, 1984.
Förstner, W., 1986: A Feature Based Correspondence
Algorithm for Image Matching. Int. Arch. of P. & RS., Vol.
XXVI, P. 3/3, Com. Ill, Rovaniemi, Finland, 1986.
Fórstner, W., 1995: Matching Strategies for Point
Transfer. In: Fritsch, Hobbie, (Eds.): Photogrammetric
Week ‘95. Herbert Wichmann Verlag, Heidelberg, 1995.
Fuchs, C., Lang, F., Fórstner, W.,1994: On the Noise and
Scale Behaviour of Relational Descriptions. Int. Arch. of
P. & RS,, Vol. XXX, P. 3/1, Com. Ill, München, 1994.
Fuchs, C., Fórstner, W., 1995: Polymorphic Grouping for
Image Segmentation. In: Grimson, E. (ed): Proc. 5^ ICCV
‘95, IEEE Computer Society Press, 1995.
Gülch, E., 1994: Erzeugung digitaler Geländemodelle
durch automatische Bildzuordnung. Dissertation at
Stuttgart University. Deutsche Geodätische Kommission,
Heft Nr. 418, München, 1994.
Heitzinger, D., 1996: 3D-Oberfláchenmodellierung mit
topologischen Grundelementen. Diploma Thesis at the
Institute of Photogrammetry and Remote Sensing at TU
Vienna, 1996.
Kager, H., 1989: ORIENT: A Universal Photogrammetric
Adjustment System. In: Gruen, Kahmen (eds.): Optical 3-
D Measurement Techniques. Herbert Wichmann Verlag,
Karlsruhe, 1989
Krzystek, P., 1995: Generation of Digital Elevation
Systems. In: Second Course in Digital Photogrammetry.
Notes of that course held at the Institute of
Photogrammetry at Bonn University, Feb. 6 - 10, 1995
Lang, F., Fórstner, W., 1995: Matching Techniques. In:
Second Course in Digital Photogrammetry. Notes of that
course held at the Institute of Photogrammetry at Bonn
University, February 6 - 10, 1995
Molnar, L., Wintner, J., Wóhrer, B.: DTM System SCOP
in a New Technological Generation. Int. Arch. of P. &
HS., Vol. XXXI, Com. IV, WG 4, Vienna, 1996
Tsingas, V., 1992: Automatisierung der Punktübertragung
in der Aerotriangulation durch mehrfache digitale Bild-
zuordnung. Dissertation at Stuttgart University. Deutsche
Geodátische Kommission, Heft Nr. 392, München, 1992
Vosselman, G., 1995: Applications of tree search
methods in digital photogrammetry. ISPRS Journal of
Photogrammetry and Remote Sensing, Vol. 50, Number
4, pp 29 - 37, Aug. 1995.
Zhang, Z., Zhang, J., Wu, X., Zhang, H., 1992: Global
Image Matching with Relaxation Method. In: Proceedings
of the International Colloquium on Photogrammetry,
Remote Sensing and GIS, Wuhan, China, 1992, pp. 175-
188.
Aeri
AB
In 1
nera
area
ma,
info
road
info
recti
The
com
vari
eval
base
inter
viga
weve
urba
imag
terp:
feati
exist
senc
Scen:
Ir
nera
scen.
imag
infor
seen
the :
map
ned
class
for t
exire
Searc
Mign
