point. It is kept fix unless the second point is iden-
tical to the first. In this case the primitive is shifted
horizontally.
e The second point chosen together with the mouse key
determines which parameters actually are changed.
e The second mouse key is reserved for horizontal —
possibly virtual — edges of the primitive and allows a
rotation of the primitive around a vertical axis through
the first point.
Table 1 shows a subsection of the association table for the
building shown in fig. 2. E. g. if point pair (2,4) is chosen,
either the form parameter b or — when pressing the second
mouse key - the form parameter b together with the rotation
can be changed by adapting point 4 to the image content.
[point]: - 10d Socles3»ob5poo* tpe
1 Ur Uy hi hi ho hi b
hi Uz Uy h2 b
2 b «a
3 hi ha ha Ug Uy ha
4 hi b b ha Ur Uy
b a
Table 2: Association table for controlling interaction
After selection of two points the parameters p; are selected,
which are to be adapted. The change of the mouse position
during moving the second point determines the size of the
parameter change. Using the difference ratios 0xp/0p: and
óby /óp; of the image coordinates with respect of the param-
eters we obtain the coordinate changes Az, and Ay, in the
case of two parameters p; and p; by
dry Sys
br) fie dm Az,
(45 - dre Se [0 D
op; op;
In case only one parameter p; is to be changed the equa-
tion system for determining the parameter is over determined.
Then we use the pseudo inverse. In this case the second point
moves on that image line which is determined by the specified
parameter change, which appears to be meaningful.
This procedure holds for all combinations of parameters de-
fined in the association table and is used for all types of
primitives as well as for the height of the prismatic models.
Changes of the association table are simple and allow a fast
adaptation to preferences of the operator.
2.4 Automation of Building Acquisition
Building extraction can use two features of Digital Pho-
togrammetric Systems for the advantage of increasing the
performance:
1. Images are digital. This allows the computer to have
access to the image content. Mensuration and clas-
sification are two basic tasks the operator needs to
264
perform, which both can be heavily supported by im-
age analysis. Mensuration is the easier task as it only
determines geometric properties of prespecified pa-
rameters.
2. DPS provide a direct and fast access to all images
in concern without repeated interior orientation. Us-
ing matching algorithms for more than two images
all information available in a photogrammetric block
can be used simultaneously, increasing efficiency and
allowing a reliable self diagnosis of the automatic pro-
cedures, based on the high redundancy available.
The difficulties in automation, mentioned in the introduction,
and which are the motivation for developping a semiauto-
matic system, will become more transparent in the following.
Using Edge Information. Image edges are lines of high con-
trast, which easily can be extracted from the image. They
carry essential geometric information. Their precision is in
the range of 1/3 to 1/20 of the pixel size, depending on the
noise level and the contrast.
Unfortunately, automatic edge extraction procedures produce
too many edges due to objects which are not of special in-
terest in the chosen context, e. g. caused by windows on
roofs, texture, shadow or vegetation. At the same time the
procedures are likely to miss important edges, due to limited
contrast or occlusions.
Therefore separating the responsibility of man and machine
seems to be appropriate: the geometric information is recov-
ered automatically, while the decisive part, namely selecting
the edges of interest is left to the operator — the procedure
used for defining the ground plan of prismatic models at eave
level.
Model to Image Matching. In case the system is provided
with a model of what to expect in the image a matching
between image and model can be performed. This matching
performs an instantiation of those values of the model which
are not fixed. These may be geometric parameters fixing the
pose or the form or class labels fixing the (sub)category the
object belongs to.
In our context the interaction provides both, the internal
structure of the model as well as approximate values for the
parameters of the primitives. The wire frame model therefore
can be projected into the image in order to find corresponden-
cies between the edges of that projection and automatically
extracted image edges. Minimizing the geometric differences
in the images in a least squares adjustment yields optimal
values for the parameters left free, either only the pose or
both, pose and form parameters.
The accuracy of the final result can be evaluated in order to
decide on further steps of the anlysis.
One-Eye Stereo and Multi Image Matching. Interaction
should take place in only one image if possible. The missing
3D-information should be automatically acquired from the
other images. This allows to use standard workstations, eases
interaction and increases acceptability of the system by non-
experts.
The before mentioned model to image matching could use
the information not only of one or two images but of several
images. In general it is recommendable to have multiple over-
lap for building acquisition just to have views of all sides of
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996
MM CN gm 0) "3 =p =p
m ON ON "r3 e+ 1
ma ~~ rr d
