nbul 2004
ally hard
asily dis-
nay dete-
age seg-
and soon
have im-
function
segment
| towards
“ven and
rticularly
es of the
ice actu-
le makes
irces that
Or orien-
y (sharp-
on tasks.
point de-
lified by
al distri-
operator
ymetrical
iC exper-
ng if the
ivourable
ted algo-
n quality
help him
the task
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part BS. Istanbul 2004
requirements. Unfortunately, this cannot be done a priori.
But as soon as one of the solutions is computed, we have
all the necessary information to display this quality indica-
tion. According to the kind of features and the number of
views that were used, one of the following measurements
is computed.
e For edges extracted from a selected segment in two
views, the angle between the two interpretation planes
is graphically displayed.
e For triangulated points, the distance between the two
interpretation lines is provided. It is expressed both in
pixel unit to reflect possible image or pairing errors,
and in world reference unit to let the user appreciate
the geometrical configuration.
e For solutions based on a single view, the maximum
deviation caused by a one-pixel perturbation applied
to each segment end is computed.
5 CONCLUSION
Manual registration of 3-D models on images remains a
difficult task. We have implemented a generic procedure
based on the 3-D interpretation of a restricted set of visual
features in order to fix two rotation parameters at least.
The registration work can then be very easily completed
through interactive means. Different possible geometric
configurations are interfaced in a quite homogeneous way,
so that this procedure provides the required geometrical as-
sistance without calling in question the high flexibility of
our interactive approach to the 3-D modelling work.
It was tested on many applications in structured environ-
ments (industry, urban sites, interior scenes), with images
or scenes with different levels of quality (low contrasts,
low illumination, highly specular surfaces), and with vari-
able knowledge on sensor calibration or scene geometry. In
all these situations, the small set of implemented solutions
could be used to bring off the modelling. Unfortunately the
ergonomics evaluations campaign that was planned could
not be held, and up to now we have only feedbacks from
some experienced users. We believe that the user inter-
face can be mastered in a very short period. But we lack
of information about the required time to acquire enough
experience to be able to select the relevant configuration
and image features. For instance a typical error is to com-
pute the vanishing point of lines that are nearly parallel to
the image plane. The numerical instability of the provided
solution causes inconsistencies that are well detected by
the quality estimation calculation, and that soon affect the
modelling if the user ignores these indications. But lot of
time may be lost if it happens too often.
This registration procedure is integrated into the new inter-
active 3-D modelling software which is still under process
in our laboratory. We plan to complete it with a sixth solu-
tion based on the interpretation of elliptical contours. This
could help in the many situations where circular shapes are
available in the displayed environment.
200
REFERENCES
Chen, H. H., 1991. Pose determination from line-to-plane
correspondences: existence condition and closed-form so-
lutions. IEEE Trans. on Pattern Analysis and Machine In-
telligence 13(6), pp. 530—541.
Chen, Z. and Huang, J.-B., 1999, A vision-based method
for the circle pose determination with a direct geometric
interpretation. IEEE Trans. on Robotics and Automation
15(6), pp. 1135-1140.
Dhome, M., Richetin, M., Lapreste, J. and Rives, G., 1989.
Determination of the attitude of 3D objects from a single
perspective view. IEEE Trans. on Pattern Analysis and Ma-
chine Intelligence 11(12), pp. 1265-1278.
Even. P., 2001. A top-down approach to interactive 3-D
modelling of architectural sites. In: XVIII” International
Symposium of CIPA, Potsdam, Germany, pp. 585-592.
Even, P., 2004. Task-oriented 3D reconstruction: rele-
vance, actuality and perspectives. In: Proc. of 3D Mod-
elling 2004, Paris, France.
Even, P. and Malavaud, A., 2000. Semi-automated edge
segment specification for an interactive modelling system
of robot environments. In: XIXth ISPRS Congress: Geoin-
formation for All, Amsterdam, NL, pp. 222-229.
Fischler, M. A. and Bolles, R. C., 1981. Random sample
consensus: a paradigm for model fitting with applications
to image analysis and automated cartography. Communi-
cations of the ACM 24(6), pp. 381—395.
Horaud, R., Conio, B., Leboulleux, O. and Lacolle, B.,
1989. An analytic solution for the perspective 4-point
problem. Computer Vision, Graphics, and Image Process-
ing 47(1), pp. 33-44.
Kanatani, K. and Liu, W., 1993. 3-D interpretation of con-
ics and orthogonality. Computer Vision, Graphics, and Im-
age Processing 58(3), pp. 286-301.
Leroux; C; Even, P., Lottin; A, Gelin, R., Idasiak, J.-M.,
Boissonneau, J.-F. and Jeanjacques, M., 2002. Flexible
3-D modelling of hostile environments with constrained
viewpoint and multiple unknowns. In: ISPRS Comm. V
Symposium: Close-Range Imaging, Long-Range Vision,
Corfu, Greece, pp. 175-180.
Quan, L., 1996. Conic reconstruction and correspondence
from two views. IEEE Trans. on Pattern Analysis and Ma-
chine Intelligence 18(2), pp. 151—160.
Shum, H.-Y., Han, M. and Szeliski, R., 1998. Interactive
construction of 3D from panoramic mosaics. In: IEEE
Conf. on Computer Vision and Pattern Recognition, Santa
Barbara, CA, pp. 427-433.
