ve
st
al
Ig
).
iC
al
ts
of
1€
is
d,
EVEN, Philippe
e C) Extraction of a line from the contours of a shape of revolution in two views or more.
This method is a generalization of the previous one. On each image, a median plane to the contours interpretation
plane is computed. The intersection of each median plane provides the 3-D shape revolution axis. Cylinders and
cones are classical shapes of revolution, but more complex shapes can also be modelled with surrounding volumes
by setting the specified 2-D lines upon the shape outline (see for example fig. 3).
e D) Extraction of a plane from the projections of coplanar features in two views.
The plane is computed from the specification of the projection of three points, one point and one line, or two lines,
according to the visible features in the image. The edge segments are interpreted the same way than in method B.
An optical ray is computed for each image point and the image projection center. Because of the inaccuracies on the
viewpoint localization and on the points specification, the optical rays of homologous points do not intersect. The
3-D point is approximated using a weighted mean.
e E) Extraction of an orientation from the projection of three orthogonal lines in one or multiple views.
This method is a generalization of an analytic solution to the interpretation of three edge segments corresponding
to perpendicular lines in space (Horaud et al., 1989). Two possible orientations are provided. The operator visually
detects the wrong one. An important requirement for a better accuracy is to avoid edge segments which interpretation
plane normal vector is nearly orthogonal to the view optical axis (grazing view).
All these methods could be enhanced if more image features were used. The information redundancy could be exploited
to achieve a better accuracy. But time is as essential requirement as precision for on-line modelling. Therefore we rather
try to reduce the amount of operator's actions.
4 SEMI-AUTOMATIC EDGE SEGMENT SPECIFICATION REQUIREMENTS
Edge segments specifications are often required when modelling with Pyramide. Manual specification is a tedious task.
Accurate drawing of a line segment over the image feature is time consuming and visually tiring. Therefore a semi-
automatic assistance has been integrated. One solution consists in first automatically extracting the edge segments in the
image, then manually selecting appropriate ones (Lang and Fórstner, 1996). But the edge segment density is sometimes
so high that the selection task requires a fine accuracy and becomes tedious. In the solution that we investigated, a line
segment is quickly drawn close to the expected edge segment. It is then automatically attracted towards the best edge
segment extracted in the vicinity of the initial solution.
Several requirements should be fulfilled to ensure the efficiency of the integrated assistance (Hsieh, 1995). The automatic
attraction should be robust, and discard disturbing edge contours that could keep the segment away from the solution. It
should also be fast enough in order to provide a good man-machine cooperation. This notion of acceptable response time
depends mostly on the operator's current activity. In the present case, the operator waits on the result to proceed on the
modelling work. Therefore the average acceptable limit is about one second.
The time devoted to the possible settings of the automatic assistance should not exceed the expected benefits. The pos-
sible failures and their recovery should not compromise the assistance over-all efficiency on the whole modelling. The
assistance should not demand a strong expertise to the operator. If the implementation requires some particular operating
mode, its learning should be counterbalanced by large benefits for the modelling task.
Figure 3: Examples of virtual edges (shape of revolution outlines) or nearly virtual edges (partly occluded segments).
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 225