equalization of the mean and the variance of the g.v.
between the two windows:
Oo: xm
—+9g, (1)
ep
g, - (g, -g,)-
gc do. g.v. of a patch pixel after and before the filter
correction;
g,.gr: mean values of g.v. of the patch and the
template;
0,07: St-d. of g.v. of the patch and the template.
In the first iterations a constraint is put on the
amplification of the contrast:
O.-
edi. Q)
9p
gmax. threshold value for contrast amplification.
Since it has been noticed in previous experiences with
poorly textured surfaces that, when the initial values are
not very good, it may happen that the g.v. variance in the
search window is very small, because just a small part of
the target shows up in the window. Therefore the
amplification factor becomes very high, affecting the
convergence of the geometric transformation to the
correct solution (basically, the scale factor may be
misjudged); after a sufficient degree of convergence is
reached, this effect disappears, so the equalization can
improve the radiometric fitting without any bias of the
solution.
Two types of targets were used on the sculpture: a set
for control points, provided with point number, and a
second one for tie points; both contain a white central
disk surrounded by a black ring (see Fig. 2). To build a
template good for both kinds of targets, the image of a
target with minimal perspective deformation has been
cut out. A synthetic copy of the target has been
generated by selecting half of a profile across the centre,
interpolating it by polynomials and rotating it around a
vertical axis, to give raise to a symmetric distribution of
the g.v. with respect to the window centre (see Fig. 1).
300 -
250-
200 -
150~
1005 bi AAA A LE
zm N ! { Lj { ;
ITE V
SFT NI i
mmm NIAE
50
Figure 1 - 3D view of the template
520
The theoretical accuracy on the image, based on the
(usually optimistic) estimate given by the inverse matrix
by |.s.m., is in the order of 0.15 mm; the R.M.S. of the
differences with respect to the same targets, measured
by the InduScan system, are nevertheless in the same
range.
4. THE SEARCH FOR HOMOLOGOUS POINTS
Once targets have been located in the images,
correspondencies must be established between image
points; moreover, control points need to be identified. If
we could approximate the object by a plane, at least in
significantly large areas, we may compute a 8-parameter
transformation, identifying manually 4 points in each
image and then transferring automatically all targets
from a reference image. In aerial blocks the flight plan
and approximate informations on the elevations may
provide with a range along an approximate epipolar line
(Liang and Heipke, 1994). The same idea may be used
with 3D objects, but the approach may fail, because it is
difficult to find good approximate orientation values for
the images: it may be necessary to extend the search for
candidates to a larger region along the epipolar line,
increasing the risk of false matchings, particularly when
the candidates all look the same, like retroreflective
targets. To avoid inconsistencies, the search area should
contain only one candidate or, in other words, epipolar
geometry should be precisely known for at least three
images. To automate the process, any available a priori
information on the imaging geometry should be
considered, but still in many cases the amount of
information may not be enough.
We opted for an intermediate solution, where human
intervention is still necessary, but limited to tasks easy
for humans and more difficult to algorithms. In practice,
we want to improve a workflow where the operator will
roughly point to the targets on the screen, then will label
them in all images, either assigning new numbers or
reading the control point numbers (see Fig. 2). This is
still much work when there are many images and many
targets, like in our case; therefore we set up a semi-
automatic procedure, where the operator's job will be
restricted to restarting the procedure whenever it stops
and to identifying the control points only. The other
targets will be numbered and recognized automatically.
Figure 2 - The two kinds of targets used
4.1 Description of the procedure
The input data for the procedure are the target image
coordinates and the object coordinates of the control
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B5. Vienna 1996
points
at lea
points
(asym
comp!
rotatic
image
onar
more
seque
values
provic
to the
orient
we |
expre:
quate
values
orient:
target
will b
points
too p:
correc
We c
geom
to fii
(Ayac
Let A
image
lines
target
are cc
