Parameter Unit Value
focal length f pix 947
aspect ratio a 1 0.94
skew s 1 0
principal point z9 pix 250
principal point yo pix 384
radialdistortionk — pix ^? —1.7-1077
Table 1: Inner camera parameters.
to photographic cameras, one pixel on the facade has a size
of approx. 3cm, whereas a disparity change of one pixel
changes the depth by about 5 cm even though camera setup
could be considered wide base stereo. A noise of 1 pixel in
the disparity map therefore results in a noise of 5cm in
the depth map which makes the surface pretty coarse. This
could be improved with an enhanced dynamic program-
ming scheme that allows subpixel disparity values. Since
this example has been computed with only three images
out of over 100, better results might be obtained if more
images are used in order to take advantage of averaging.
To verify the influence of the relief on the visual im-
pression of the results, three positions have been marked
in Fig. 5. For the flat model it can be seen that the rela-
tive position of the marked features does not change with
respect to their surroundings. Both views are related by a
planar homography.
For the relief model, different angles of view lead to dif-
ferent visibility of details that are below or above the fa-
cade. Bell and sky (a) can only be seen from the right
view. The statue (b) which is embossed on the facade
changes its relative position to features directly on the fa-
cade. The door frame blocks off the view onto the leftmost
part of the door in the left image, making the sunlit part (c)
smaller than compared to Fig. 5.
There are some gross errors in reconstruction of the facade.
Fig. 9 shows the left bottom part of the model where errors
during guided matching lead to a cavity. This and other ar-
rors occur probably because the epipolar lines are parallel
to the horizontal structures on the facade due to horizontal
movement. In areas with little or even irritating texture this
will mislead the matching. One possibility to circumvent
this effect would be to use images taken from a different
height so that the epipolar lines run diagonally or vertically
across the facade. Errors like these however do not influ-
ence the outline of the reconstructed model because depth
is forced to given values at the borders.
5 CONCLUSIONS
A hybrid model that refines a coarse wire frame model by
detailed relief recovered from images has been proposed.
The approach is suitable for rapid prototyping, because the
required model can easily be constructed by manual inter-
action whereas the relief will be generated automatically.
Results including a comparison with a flat model are given
for a set of images taken by a hand held video camera.
972
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Figure 9: Gross error in reconstruction. This figure shows
the lower left part of the facade.
They show that relief structure can be retrieved and that re-
lief information can upgrade visual impression. The over-
all quality is limited by the resolution of the camera but
improvements are expected if the number of images is in-
creased. Additionally, the dynamic programming scheme
could be enhanced to allow a disparity estimation with sub-
pixel accuracy.
REFERENCES
Baillard, C., Schmid, C., Zisserman, A. and A.Fitzgibbon
1999. Automatic line matching and 3d reconstruction from
multiple views. In: ISPRS Conference on Automatic Ex-
traction of GIS Objects from Digital Imagery, Vol. 32.
,
Brenner, C., Haala, N. and Fritsch, D., 2001. Towards fully
automated 3d city model generation. In: E. Baltsavias,
A. Grün and L. van Gool (eds), Proc. 3rd Int. Workshop on
Automatic Extraction of Man-Made Objects from Aerial
and Space Images.
Falkenhagen, L., 1997. Hierarchical block-based dispar-
ity estimation considering neighbourhood constraints. In:
Proc. Int. Workshop on SNHC and 3D Imaging, Rhodes,
Greece.
Grün, A., Remondino, F. and Zhang, L., 2003. Auto-
mated modelling of the great buddha statue in bamiyan,
afghanistan. In: Photogrammetric Image Analysis, IAPR,
Vol. XXXIV-3/W8.
Hartley, R. and Zisserman, A., 2000. Multiple view geom-
etry in computer vision. Cambridge University Press.
Koch, R., 1997. Automatische Oberfláchenmodellierung
starrer dreidimensionaler Objekte aus Stereoskopischen
Rundumansichten. PhD thesis, Universität Hannover.
Pollefeys, M., 1999. Self-calibration and Metric 3D Re-
construction from Uncalibrated Image Sequences. PhD
thesis, Katholieke Universiteit Leuven.
Teller, S., 1998. Automated urban model acquisition:
Project rationale and status. In: DARPA Image Under-
standing Workshop.
KEY
ABS]
In thi
doma
scene
(i)Ve
wiref
vehic
ribbo
result
vehic
other
tempi
is per
under
veloc
SAR
Parti
deter
layoy
simul
Traff
Rese:
appli
and .
estim
Furtl
on c
autor
guid:
emer
infor
situa
wher
appli
and
Syst
The
the E
and
perf(
indu
S