International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
configuration
number of points
5 [ qe 15
"Tetra’: camera distance approx. 3 m 83% | 28% | 09%
"Airl: camera distance approx. 1500 m 21594 41.7955 | 80.6 96
'Air2: camera distance approx. 1500 m 5 90 7% | 06%
number of points
configuration T.
Bay » ^35
'"Strect1': camera distance approx. 10 m 50 96 50 96 25 96
'Street2': reference distance approx. 2.85 m 50 96 25 % | 125%
Table 1: Minimum thickness in percent of the camera distance for which the computation of the trifocal tensor was still
successful. For the configurations Tetra’, ’Airl’ and ’Air2’ the given values hold for any computation method, whereas
for the two ’Street’ configurations they hold only for the constrained method minimizing reprojection error (CR).
k points, (ii) retrieve the exterior orientation parameters
for the relative orientation and since the camera is cali-
brated use the known calibration (2061.0, -1339.0, 2292.8)
for this purpose. The extracted orientation parameters
shall serve as approximate values for a subsequent bundle
bock adjustment of the three images (with fixed interior
orientation and fixed distortion) using the subset of the k
points. After the block adjustment all 121 points in the
three images are computed in object space by spatial in-
tersection. This Euclidian reconstruction differs from the
known positions of the retro reflecting targets by an ab-
solute orientation A, i.e. shift, rotation and scale. This
transformation A is computed using all 121 points. The
remaining discrepancies after the absolute orientation are
used to judge the quality of the initialization of the block
adjustment.
The aim is to find the smallest possible subset of points
for the computation of the tensor, which still provides
good enough approximate values for the initialization of
the bundle block adjustment. This task can already be
solved with the minimum number of 7 points, which are
shown in figure 3 right part. And as it turned out, for
the creation of the approximate values it is not relevant
whether the tensor is computed in the simple way "UCA’
(without the constraints and minimizing algebraic error)
or in the rigorous way ‘CR’ (with the constraints and min-
imizing reprojection error). The remaining errors after
finishing the bundle block adjustment, using the known
calibrated interior orientation and the known non-linear
distortion parameters, and the absolute orientation are:
reconstruction errors [m]
mean max
x. 0.009 -0.030
y | 0.009 -0.033
z | 0.006 0.030
For another task with the real data we could also ne-
glect the known interior orientation of the camera and use
Kruppa's equations to derive a common interior orienta-
tion for the three images. In this case, however, the depth
range of the used subset of object points has to be ex-
panded a lot, see figure 3 right part, and at least 15 points
must be used: 11 points from within the facade, 2 points
from the roof (lying 3.3 m behind the facade) and 2 points
3.1 m in front of it. For this point sample the rigorous
computation of the tensor in the Gauss-Helmert model
by minimizing reprojection error and considering the in-
ternal constraints (method 'CR/) is necessary, because for
the direct linear solution (method 'UCA") no valid inte-
rior orientation can be obtained using Kruppa's equations.
The remaining errors after finishing the bundle block ad-
justment, using the determined interior orientation (2059.7,
-1043.8, 2726.6) fixed and without any non-linear distortion
parameters, and the absolute orientation are:
reconstruction errors [m]
mean max
x1 3115 0.326
vii 0.113 0.370
z | 0:100 -0.429
5 SUMMARY
Concerning the computation of the trifocal tensor we dis-
covered the following using different synthetic examples:
e the difference between minimizing algebraic and re-
projection error in the computation is negligible the
more point correspondences are used and the more
the respective object points deviate from a common
plane,
e minimization of reprojection error is more important
than considering the internal constraints
e if the image geometry is not too bad and at least 10
point correspondences are used, a minimum thickness
of the object points of about 1 % of the camera dis-
tance is already enough to allow a proper solution for
the trifocal tensor.
International Arch
JHCTRIHCHUT 13 0H
Radial
Canon
207
15 ——
-B5 de
Distortion [pixel]
-10 A——
-15 d
Radial
Figure 3: Left ;
Right part: Th
mark the points
on the roof signi,
the trifocal tensc
Guided by this
using real image:
we see:
if the interio
the exterior
sor is sufficic
even in the
and even if 1
number of "
mum thickn
1 96 of the
away from t
e ifthe interio
computed w
icantly away
values for th
ing Kruppa’
Although for mo
ple direct linear
rigorous solution
turn similar resul
one, because for
to yield a usable
Therefore the re
tion is to first est
Gauss-Helmert n
mon interior oriei
and finally to ini
à bundle block a
by additionally n