human operator interprets the scene by modelling a
geometric approximation of it in the CAAD program
(figure 3). The measurement is then handled
automatically by the DIPS, based on this
approximation. This procedure is referred to as 3D
feature extraction.
1.3 CAD-based 3D feature extraction
In order to establish feature extraction that provides
for high precision as well as for reliability, a top-
down strategy is chosen. The semantic object model
is used to detect the features described by this
model. Thus only relevant features are extracted
and redundant information and data complexity are
reduced to a minimum.
The three-dimensional position of the object is
derived by simultaneous multi-frame feature
extraction, whereby the object model is
reconstructed and used to triangulate the object
points from corresponding image points.
It is evident that in most cases linear boundaries
(edges) of an architectural feature contain more
information than the vertices (corners) of this
feature. Although edges are only a small percentage
of the whole image content, they have major
importance for the description of object
discontinuities. The CAD-based 3D feature
extraction routine takes advantage of this
knowledge. It first locates the edges of the features
to be measured and then derives the vertices as
intersections of appropriate lines.
The position of the edge is determined with
subpixel precision by fitting a second-order
polynomial in the direction of the gradient. The
maximum point of the fitting curve corresponds to
the subpixel position of the edge. The covariance
matrix of the estimated polynomial parameters
represents the accuracy of the edge point.
The 3D feature-extraction is described in more detail
in [4][5].
1.4 Automation
We described the 3D feature extraction as a semi-
automated top-down procedure. A CAD-generated
feature is matched with corresponding images. In
fact, any computer-vision strategy has ultimately a
very strong top-down component. Theorists have
pointed out that this is true for human perception
as well [7][8]. Seeing is largely recognizing, is so to
speak a top-down much more than a bottom-up
process.
Human perception happens simultaneously at many
different levels. If we want to, we can perceive the
world around us as being composed of lines or of
colours. But it is most natural for us to see it as being
composed of objects. It's at the level of objects that
238
we can understand the world. To lift the degree of
automation to a higher level, we argue that it is
necessary that also in architectural photogrammetry,
the evaluation should be based on the notion of
objects.
This is true not only for the providing of qualitative
guidance in the computer measurement process,
which we will discuss later on. The notion of objects
is also the prerequisit for a possible interpretation of
the scene by a computer-program. It should be
mentioned here, however, that an automatic
interpretation of an architectural object faces many
difficulties, not the least of them being that there's
just simply no one correct way to model any
architectural object. This is obvious not only from
well-known texts about architectural theory
[9][10][11]. It also quite simply has to do with the
fact that no two CAAD-operators will model the
same building the same way.
This doesn't mean that automation has to stop here.
Increasing automation of the whole modelling and
measuring process could instead be achieved by a
computer-learning mechanism, that allows the user
to teach his modelling preferences to the system.
An object-oriented data-integration is, as we will
show, the essential prerequisit for these
functionalities.
2. DATASTRUCTURES IN DIPS AND IN CAAD
2.1 General Considerations
= —_—"
d uniquepointe S |
[3DObject —
ocho
Figure 1: Schematic comparison of Datastructure in DIPS and CAAD:
uniqueness of points is not adhered to in CAAD
Comparing the datastructures that are commonly
used in digital photogrammetry and the ones found
in CAAD systems, one essential difference can be
stated as universal. It is the use of unique points in
photogrammetry that is not adhered to in CAAD.
CAAD datastructures are geared towards modelling
capabilities, for which discrete elements have
proven to be useful. (see figure 1).
Furthermore it can be said that structuring means
that go beyond points, lines and layers are to this
day rather rare in photogrammetry systems. Means
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
of higher 1
in CAAD,
photogram
Another di
for the ex«
involved i
usually not
no standa
whereas fo
is used in 1
world: DXI
While ther
improve c
exchange
pointed ou
in discussi
developme
and CAAD
structure a
how a dat:
and pro
informatio:
2.2 Data-S
System
The Datas
project ca.
There is o:
that refere
(unique m
identical :
representa
every imag
the 2D po
but they :
point or ol
3D repres
extraction
For the
Structure n
2.3 Data-S
Standard
DXF is on
formats ir
back to 19
AutoCAD,
between c
operating
and chang
AutoCAD.
software,
almost a
market. Sc
facto stanc
