Shih-Hong Chio
nature in the object space could be part of a roof edge. Among all the meaningful line segments the system chooses the
best one and prompts it on the screen with an initial guess of the possible form of that roof patch. The operator then has
to confirm it or deny it interactively on the screen.
If the operator denies it, the computer tries to find another one and prompt it again on the screen. Since human operator
is good at the interpretation and the computer is good at the computation, we let the operator do the examination of the
results computed by the computer. Immediately after each action of the operator, the computer starts to recalculate the
new situation provided by the new information from the operator's action. It prompts new suggestions on the screen.
Eventually an initial line segment will be accepted by the operator. In the following we will call this first confirmed 3-D
line the initial 3-D line. This line in general is only a segment of a roof edge.
After confirming the initial 3-D line, the operator then has to select relevant line segments (mostly also only broken
fragments) associated with the initial 3-D line. By association we mean that they, together with the initial 3-D line,
seem to form a roof patch. In the following we will call these relevant line segments the associated lines. The associated
lines could be 2-D or 3-D, depending on how they are linked and matched in the first stage of finding meaningful lines.
The selection is done mono-scopically in one image, therefore it can not be hundred percent correct. If the new
information is enough, the system tries to reconstruct the entire roof patch. Otherwise the system will ask the operator to
point the cursor to the approximate location of the roof corners. The system will find the best point from the Point
Database for that corner. After that, the system starts to compute the 3-D coordinates of all the corners. The roof patch
is then reconstructed from the corners. The result is displayed either in perspective view on the screen or
stereoscopically in two images. The operator can examine the result. If the operator is not satisfied with the result, he
can modify it interactively. If for any roof patch the finding of the initial 3-D line failed at the very beginning, that roof
patch must be reconstructed manually by the operator by indicating the corners and assigning a roof patch model to it.
Thus, the verification of the initial 3-D linear segment, the selection of relevant 2-D or 3-D line segments, the pointing
to the missing corner(s) and the modification of results are the main interactive components, which characterize the
whole system.
2 SYSTEM DESIGN AND ILLUSTRATION
^ Fig. 1 shows the diagram of this interactive
es ; ; :
system for roof patch reconstruction. All the
Semantic Information .
necessary lines, no matter 2-D or 3-D, are all
prepared in a forgoing program. In that program
straight lines are constructed by simultaneously
Another 3-D
Line No \ Boundary of Roof Patch
Yes
Verification and Correction of
Semantic Information linking and matching of consecutive single edge
Ÿ . E . . . ^ 3
Selection of Relevant 2-D and 3-D Linear Segments pixels m the two stereo Images under the
in Single Image consideration of the general knowledge about
Yes roof edges in the object space. Here, the general
—» Position of Lacking Corners in Single Image 4 knowledge we assumed is that a roof edge is
1 limited to a straight line in the space. It is either
No . . - Pa 7 C
Determine the 3-D Coordinates of Corners Acoording to horizontal or oblique. Thus, curvilinear roof
Point DataBase Combination of Three Different Approaches edges are not considered here. With this
Y
knowledge in mind, we can link and match edge
pixels, extracted by low-level feature extraction
methods, more effectively into meaningful line
segments. Details about the method are described
in a previous publications [Chio et al., 1999].
At the beginning of the roof patch reconstruction,
the system will display all successfully linked
and matched 2-D and 3-D lines in both images of
the stereo pair. For each roof patch, one of the 3-
Fig.1: Diagram for roof patch reconstruction D lines will be chosen by the system as the initial
3-D line for starting the reconstruction process. That line will be prompted on the screen with different color. The
system will then select a roof patch model associated to this line and prompt the semantic information of that roof patch
model on the screen.
Only two kinds of roof patch models shown in Fig.2 are assumed in this research. They represent the most encountered
cases in Taiwan. Here we suppose that a roof patch is either quadrangular slope or flat plane in the object space. Note
that one complete roof might consist of more than one roof patch. For example, a gable roof with one single ridge is
composed of two slope roof patches. The selection of the roof patch model depends firstly on the building construction
and secondly on the generalization by the operator. By assuming only these two simple models, the semantic
information of 3-D linear segments becomes very simple: namely for slope roof patch either oblique or horizontal
Inference the Roof Patch on the Basis of
the Model of Roof Patch
184 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.