were detected and reconstructed (cf. Figure 3). The second
strategy which relies on a segmentation of the range data set
only gives correct results for the flat data set, in the suburb
data, the segmentation is not strong enough to feed the para-
2.5 Reda Fayek
Data Source: Range data, data set flat
uildings are de-
ted properly
terms of the un-
struction phase.
> data, data sets
| on the generic
ar then their sur-
ption parametric
ar buildings with
e buildings is 10
| of stereo-image
/0 steps, namely
] the reconstruc-
f each detected
gions of interest
er than their sur-
ain height differ-
ig reconstruction
stereo image. In
matched to form
proximate paral-
rametric building
es. The building
s chosen to rep-
proach aims at a
tracting 3D-lines
J the parametric
ala 1995] and in
on the flat and
results could be
19s could be de-
iburb 30 from 38
metric model correctly (cf. also Weidner).
Figure 3: Perspective view of the detected and reconstructed
buildings in the data set suburb
2.4 Klamer Schutte
Data Source: Single image, data set flat
Object Model: Buildings are modelled as polyhedral ob-
jects, what results in a polygonal description, in terms of the
projected faces. For each object in the database a set of as-
pects is generated. Depending on the light source shadow
regions are defined. Each aspect is defined by a collection of
constraints concerning regions, contours, corners and their
relations.
Prior Knowledge:
> Camera position and parameters,
> light position and direction,
> approximate size of buildings (xsize between 5 and 14
m, ysize between 10 and 30 m, h1 between 1 and 20
m, h2 between 1 and 10 m),
> noise in images (gaussian noise with o = 10),
> generic parametric object descriptions which are trans-
formed into aspects interactively.
Strategy: Object hypothesis are generated in a relaxation
step, where the regions in the segmented image are matched
to the aspects from the data base. This step is followed by an
estimation and verification process, to exactly fit the objects
to the image data. A detailed description of the approach is
given in [Schutte 1994].
Results: Not all buildings present in the image are found.
This is mainly caused by errors in the segmentation of the
image, which cannot be recovered. Another cause is that the
models used are quite simple (i.e. not including windows is
the roof.) However, even if these models were incorporated,
the segmentation of the image for such buildings proves to
be extremely difficult.
For some buildings, multiple hypotheses are found. This is
a ‘feature’ of the system used. The reason is that some-
times hypotheses are generated for 3 segments found (2 roof
+ shadow), and sometimes for less.
Some buildings found do not match exactly. This on one hand
due to segmentation errors. Missed or incorrect boundaries
Will result in wrong object parameters. On the other hand it
is due to the fact that not enough segments are found. If, for
example, the shadow is not found, it is impossible to estimate
the height h1 of the house. Since only one image is used, the
Z parameter is not very accurate.
771
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Object Model: This approach does not base on an ex-
plicit building model, but on the idea, that man-made ob-
jects consist of planar faces and are of a certain size. For
a detailed description see e.g. [Fayek & Wong 1994]. The
generic building model consists of one or more vertical (or
near-vertical) surface patches, together with the neighboring
surface patches. Groupings of these patches have to have a
certain size.
Prior Knowledge: This approach starts with a triangular
mesh of the raw range data. The general strategy is to iter-
atively coarsen the given mesh while preserving topographic
details of the original data. The mesh coarsening requires
certain preset parameters controlling the allowable surface
approximation errors.
> only surface patches larger than a certain minimal size
are allowed to initiate a possible man-made surface en-
try. The threshold is set to 5 m?.
> A potential building has to consist of a collection of
patches with a given minimal size. This size threshold
is set to 300 m?.
> Man-made structures are separated from the back-
ground by the simple strategy of being enclosed by
a larger nearly horizontal patch corresponding to the
background.
Strategy:
> Triangulation of the range map.
> Segmentation of range map into nearly planar patches.
> Categorization of patches according to their slope.
> Instantiation of suitable patches for man-made struc-
tures.
> Growing of initiated models and final validation.
Results: The procedure principally aims at a detection and
recognition of the man-made structures. Thus the result
proves a good localization of the houses, hovever only a
rather poor reconstruction.
The results on the data set flat are given in terms of the 3D-
coordinates of the center of gravity of the building, the num-
ber of nearly-planar patches it consists of, the total outside
surface of the building, and the total surface area covered by
the it. 13 buildings could be detected properly, 4 buildings
were reconstructed as being aggregated in 2 pairs. The au-
thor points out that for fast recognition even a coarser mesh
is sufficient.
Figure 4: Detected buildings
