nbul 2004
r left tri-
id the or-
onstraint
t (Koch,
its given
nage be-
iis limits
els to no
lecessors
(6)
ems with
th neigh-
etely un-
ong path
‘een con-
d to keep
receding
(7)
(8)
>and 2-1
; a small
'ighbour-
iitial val-
le.
map for
input im-
ilculated.
y triangu-
}.10.2.30
on of the
is results
Figure 5: Two artificial views onto the west facade of Alex-
iuskapelle with flat surface model. See text for explanation
of details: (a) sky, (b) statue, (c) shadow.
in a set of elevation grids which can be combined into
one final elevation grid by means of planar homographies.
Depths that belong to the same zy-position will be over-
layed by these transforms. Some robustness is gained in
this way, because each depth has been calculated indepen-
dently — with respect to errors of the guided matching —
and the final depth is computed as an average or median
depth.
2.6 Texturing the Relief
Similar to the depth, the appropriate color is taken from
one input image from the respective location. In principle,
color could be taken from all images but then it has to be
solved which pixels are adequate. The final result is a reg-
ular 2D raster where each cell contains a depth value and
an associated surface color.
3 EXPERIMENTS
Test Data Tests have been carried out on test data of a
chapel (west facade of Alexiuskapelle, Ettlingen). A video
sequence had been taken using a SONY DCR VX-2000
handheld video camera while walking along the pavement
across from the chapel. The camera had been pointed to-
wards the facade so that the images are all slightly con-
vergent. The image format is standard video resolution of
720x576 with the camera held sideways and the zoom set
to the minimum focal length of about 6 mm (1/3" CCD)
in order to fit best to the dimensions of the facade. From
the whole sequence, three images — each about 100 frames
apart — have been selected for processing (see Fig. 3).
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Figure 6: The same two views as in Fig. 5 but this time with
recovered relief taken into account.
Camera Calibration Because the camera is a standard
consumer product originally not intended for measurement
purposes, its inner parameters were unknown and had to
be determined by means of self-calibration. Control points
have been chosen manually to eliminate possible outliers
that could reduce the quality of the reference frame. The re-
sults of the bundle adjustment for the inner parameters are
given in Tab. 1.
Along with the inner parameters of the camera, a metric
reference frame for all points has been computed. An ab-
solute scale has been introduced by setting the width of the
fassade to 6 m. The 3D-coordinates of the control points
have got a standard deviation of 14 cm.
Rectification and relief recovery An example for a rec-
tified image pair is given in Fig. 4. One can verify that
corresponding points indeed lie on the same image row.
Two different models have been generated according to
the processing chain as described in the previous sections.
The one in Fig. 5 assumes a flat surface whereas the other
is shown with the recovered relief (Fig. 6). In order to
demonstrate the differences, both models are shown from
the same two artificial viewpoints.
4 DISCUSSION
On first impression, the quality of the flat model looks bet-
ter because straight lines look much smoother than in the
relief model. This is due to geometrical errors of the relief.
Because the resolution of the video camera is low copared
