image(s) [1
;
segmentation
recognition:
database object detection
folies | — LA
models reconstructed
labelling object
Figure 1: Object reconstruction in image understanding.
and imaging geometry can be selected to optimize the pho-
tography for a given application. Image scales in the order of
1: 4000 to 1: 10 000 are commonly used. Nadir photography
with forward overlap of 60% is practical for most urban map-
ping applications. Multi-photo geometries are advantageous
in providing redundant information and improving the accur-
acy of reconstruction. Colour photography is preferred over
panchromatic photography not only because of the additional
segmentation cue, but also because image interpretation is im-
proved. Similarly, stereo imagery is preferred over monocular.
False-colour infrared photography is useful in automating the
separation of man-made object, vegetation and water bodies
by image segmentation. As a rule of thumb, photography
should be scanned to achieve a minimum spatial resolution
of ca. 25cm. This is generally a good compromise between
managable image size and detail for many applications.
The measurable features in near-nadir photography are
primarily building roofs and the structures on them. Walls,
while potentially visible when a building appears at other than
the nadir point in a photograph, are commonly obscured by
eaves or by the camera viewpoint, subject to shadows and
imaged with poor geometry such that they cannot be reli-
ably measured. If the building model should include accurate
wall detail, terrestrial measurements, e.g. photogrammetric,
become necessary. Alternatively, building walls may be in-
ferred by projection of the reconstructed roof perimeter onto
an underlying terrain model. Eaves may also be modeled in
the reconstruction if their dimension is known (Mason and
Streilein, 1996).
Other data sources for building reconstruction include digit-
ized maps and plans (Nebiker and Carosio, 1995). Technolo-
gical developments promise im proved data sources in the near
future, such as direct digital aerial imagery (Hofman et al,
1993) which avoid the need for scanning, and laser-scanners
(Krabill, 1989) which provide direct 3D coordinates at higher
accuracy (but currently lower spatial resolution) than stereo
techniques
2.3 Fidelity of Reconstructed Models
Quantifying model fidelity is a prerequisite to judging the util-
ity of reconstructed buildings for an intended application. It
consists of two parts: the geometric accuracy of the recon-
struction and model completeness. Model completeness must
be judged against the level of generalisation required by the
application. In order to quantify geometric accuracy, the ac-
curacy of all components in the reconstruction process need
to be accounted for and propagated as uncertainties in the re-
construction. This requires knowledge of camera calibration,
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
the effects of photograph scanning, etc. Importantly, accurate
camera models, exterior orientation information etc., can be
exploited in feature extraction, e.g. in using the epipolar con-
straint and providing a statistical basis for generating hypo-
thesis in matching and for fusing different information sources
(see Forstner, 1993; McKeown and McGlone, 1994).
2.4 Semi-automation as Goal
Building reconstruction is difficult. Buildings themselves are
complex, often irregular, structures, varying in appearance
according to function, cultural, climatic and topological con-
text, building codes, architectural and aesthetic tastes, con-
struction materials, colour, state of repair, etc. Their ex-
traction from imagery is difficult due to shadows, weak con-
strast, noisy imagery, interferences such as trees, loss of in-
formation due to perspective projection, occlusions and so
on. Nevertheless, humans are generally very effective in cop-
ing by being able to “understand” the imagery using a vast
skill /knowledge-base. The key to automating image under-
standing is, therefore, to give a computer the background
of common-sense knowledge and skills that humans possess.
Practical results in coding and using common-sense know-
ledge in Al have, however, not been achieved nor are forsee-
able (Dreyfus, 1994). This problem is fundamental; we do
not yet understand how the human performs such task und
thusfar have developed ad hoc methods (Fórstner, 1993). Im-
age understanding systems therefore require integration of the
superior interpretative skills of the human operator to achieve
operational performance. It is this fact which motivates the
work described here.
In an operational context, semi-automated approaches to ob-
ject reconstruction are justifiable when they are able to deliver
results in less time and/or more reliably than a user with con-
ventional (manual) tools with at least com parable accuracy. It
is also desirable that the expertise requirements of the user be
reduced. To this end, we expect that most gains can be made
by relieving the user of the measurementtask. Accurate image
measurement using conventional technology demands expert-
ise, is time-consuming, tiring and consequently, not without
errors. Interpretation, whilst still requiring experience, comes
more naturally to the untrained person.
3 OBJECT MODELING FOR BUILDING
RECONSTRUCTION
The most important user input in a semi-automated object
reconstruction system is selection of an object model. Ob-
ject models may contain geometrical, physical, functional and
other elements and must be view-invariant in order to be in-
variant to the observation process used to infer the presence,
form, class, etc. of object instances. We focus here on the
modeling of building geometry as the most common objective,
although other properties such as radiometry (including col-
our), texture, etc. are often important (cf. Forstner, 1993).
The representational criteria include: (i) 3D modeling cap-
ability, (ii) generic, i.e. capable of modeling most, if not
all, buildings including those complex and irregular, and (iii)
compatibility with a user's intuitive representation of buildings
observed in imagery. Following a short review of object rep-
resentations used in building reconstruction, a new modeling
scheme was developed which fulfills these criteria.
A comprehensive overview of object shape modeling is given
in Braun et al (1995). Parametrized volumetric primitives
518
(i.e. fi
recons
1995;
providi
lar or :
databa
buildir
Lin et.
gular
model
closed
also b
face n
repres:
constr
model
resent
are in:
roof d
CAD :
repres
which
els of
to rec
trol hc
and g
buildii
charac
and H
ial im:
(rectli
tensit:
Moha
(1995
constr
comm
the sc
combi
we tel
for im
systen
been :
struct
aspec!
3.1
CPS.:
