converted to solid representations, as may be required by
some applications; (b) large number of surfaces may be need
to be reconstruct complex buildings; and (c) surfaces which
are not visible in - or reconstructable from - the imagery, e.g.
occluded walls, need to be inferred to obtain a complete volu-
metric model. Note that these problems are partially resolved
when modeling by combinations of CSG primitives (Braun
et al, 1995). We believe, however, that the CPS modeling
scheme is more general. Because buildings are modeled at
a lower level, some potentially ambiguous situations can be
avoided, as might occue when only one surface of a building
component is visible. Moreover, eaves and similar structures
are more easily modelled by surfaces. Of equal importance,
we expect that surface primitives are more intuitive to the
user and that their correct reconstruction is easier to validate
than CSG primitives. In the following section an implementa-
tional scenario for CPS modeling in a semi-automatic building
reconstruction system is described.
4 TOWARDS AN OPERATIONAL
RECONSTRUCTION STRATEGY
4.1 Modes of Interaction
Two basic paradigms of user-system interfacing for semi-
automatic object reconstruction can be identified: (1) cor-
rective, in which fully automatic reconstruction is attempted
(see Fig. 1) with any errors being corrected a posterior using
manual methods; and (2) instructive, in which the computer
Is conveyed what to extract by user instruction of the appro-
priate object model and possibly identification of the object's
image location. As with (1), errors by automatic methods
must be corrected manually.
In the absence of models for reliable general scene understand-
ing (see Sec. 2.4) the corrective paradigm, however desirable
from an Al point of view, remains unoperational. Firstly,
correcting large numbers of errors is generally more time con-
suming and difficult than manual methods. The user must
intrepret the scene and also understand the interpretation re-
constructed by the computer. Then, detected errors must
be manually corrected. Secondly, the danger is greater that
some errors will be missed than errors in the case of instruct-
ive reconstruction. Thirdly, the computational expense of the
necessarily complex procedures for automatic reconstruction
may well render them slower than manual or instructive meth-
ods.
Instructive reconstruction can take different forms (cf. (Lang
and Schickler, 1993; Haala and Hahn, 1995). Our philosophy
is to interact minimally but early, assisting automated pro-
cesses by supplying sufficient information to ensure a reliable
result and in so doing limiting corrective action. The degree
of assistance will depend on the robustness of the automated
processes and the extent to which their failure can be self-
diagnosed and reported back to the user. The CPS modeling
scheme is thus well-suited to implementation in an operational
environment.
4.2 Interaction Model for Reconstruction based on CPS
Modeling
An interaction model for building reconstruction based on
modeling object using the CPS schema is depicted in Fig. 4.
The basic interactive steps are described. Potential for auto-
mation of some of these steps is described in Sec. 5. The
first interactive task entails identifying a buildings to be re-
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
decomposition, wherein the user decides on the appropriate
combination of surface primitives from the database in Fig. 3
for modeling the building. Each of these primitives is con-
veyed to the computer with the approximate location of the
surface to guide the extraction step (detailed below). The
result(s) are presented to the user for verification, possibly
in stereo-viewing mode. Once all component surfaces have
been successfully reconstructed, a composition process con-
nects these surfaces, exploiting neighbourhood relations and
possibly user interaction, to construct the building model.
Due to occlusions etc. this model may not be a complete
3D reconstruction. A final step is need to infer, e.g. walls
and missing surfaces, to form a complete volumetric model.
A strategy for the automated surface reconstruction step is
explained in Sec. 4.3.
image(s)
p identify __p »- select
building decamipose surface
select next Sect
— Surface
surface model
no object
extraction
COMPOSE 3% complete a2 verification
mode ;
reconstructed not o
object
6. 7 ; interactive
2 infer correction/
-41—— volumetric |
a model extraction
Figure 4: Reconstruction strategy based on CPS modeling.
Building model selection entails two basic interactions: (1)
indicating the approximate location of the surface; and (2)
selection of the relevant geometric model. Options for task
(1) include pointing to the surface in an image in 2D or 3D, or
marking a box encompassing the surface in an image. These
are simple “mouse-clicks” and do not require fine-pointing.
Geometric model selection (2) is made from the surface prim-
itive database illustrated in Fig. 2, and should reflect what
the user understands is extractable from the imagery. Diffi-
cult scenes will require Class 1 primitives; composite mod-
els can be used on simplier scenes. This database is small
and thus well-suited to display as graphical icons for intuit-
ive understanding and selection. The database could also be
customized to suit a specific domain (see Sec. 4.4). We are
currently exploring the potential and feasibility that additional
information conveyed by the user, e.g. surface type (e.g., wall,
roof, window), building function (e.g., house, factory, school)
and the context (e.g., light industrial, medium urban, CBD),
might have in semi-automatic building reconstruction.
4.3 Strategy for Surface Extraction
The main task of the computer in our model of semi-automatic
building reconstruction is the precise and reliable extraction
of 3D surfaces corresponding to the primitive surface model
selected in a designated image window. The strategy illus-
trated in Fig. 5 is proposed. The key steps are outlined below.
Note that the images and the object model are inputs to each
step.
520
constructed. This is followed by a mental process of building
(a)
Figure
(b) ex
in the
the D:
