Fig.6: Red channel of the right image.
As mentioned before, a transformation of image grey
values has to take place in order to reduce correlation
between observations. It is understood that this should
not lead to a loss of information, i.e. that the original
grey values have to be recoverable from the transformed
ones. Besides the fact that a loss of information is
generally not desirable the possibility of computing
object grey values for orthoimage generation has to be
maintained. One possible transformation is the
computation of differences between the channels. If one
original channel is used as input for FAST Vision, the
differences of the other two channels with the former one
can be used instead of the original channels. For the
experiments described in this paper these images were
used as input in combination with the green channel.
The results of this surface reconstruction using FAST
Vision for multi-channel images were compared to those
results obtained from using grey value images as input.
The purpose of these experiments was not to compare the
surface reconstruction with surface values derived by a
different method. Besides the fact, that the first method
yielded colour orthoimages, the differences of surface
reconstruction were not significant. This can be attributed
to the fact, that the grey value pictures already contain
the necessary information.
Fig. 7 and 8 show an enlarged part (the area entirely
covered by vegetation) of the green channels of the left
and the right image. This area is displayed here because
of the great differences of the three colour channels,
which occur especially here, whereas there are good
contrasts in all other parts of the three channels. White
crosses indicate the position of the grid points of the Z-
facets in image space. It is obvious that there are some
ambiguities in the colour values which have to be
overcome by regularization. Unfortunately, the
ambiguities are present in the colour as well as in the
grey value pictures.
Each of the 16 surface heights depicted in Fig. 7 and 8
was reconstructed by FAST Vision as the centre of a
window containing 8-8 Z-facets. These in turn contained
4-4 colour value facets each (4-4 grey value facets resp.).
The colour value facets had a size of 4m in object space,
thus the Z-facets had a size of 16m-16m in object space.
Each colour value facet contained approximately 4-4
pixel. So the number of observation equations in each Z-
facet was approximately 250 in each grey value image
and 750 in each colour image. The number of unknowns
in each of the windows containing 8-8 Z-facets, i.e. the
window from which one of the surface heights in Fig. 7
963
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
and 8 is derived, amounts to 1172 using grey value
images as input and 3354 using colour images.
The mean standard deviation of surface heights, i.e. of
the unknowns Z of FAST Vision, was 0.2m. This
corresponds to less than 0.2%, of the flying altitude of
1200m from which the pictures were taken. The
differences between Z-values derived from colour images
and those derived from grey value images was below
0.1m in all cases.
Fig.7: Reconstructed surface heights in the right image
(green channel).
Fig.8: Reconstructed surface height in the left image
(green channel).
5. Conclusion
This paper shows the latest modification of FAST
Vision: the use of vectors as observations instead of
scalars. By taking this step the full exploitation of image
information becomes possible, i.e. the use of all channels
of multi-channel imagery as input for surface
reconstruction. The progress in data processing
technology allows to handle the increase of data to be
stored and processed which multiplies with the number
of image channels being used as input. The experiments
shown here using colour aerial pictures as input result in
a reconstructed surface which is not significantly
different from that reconstructed from grey value images.
Further tests with all kinds of image material will
probably show cases where the input of multi-channel
imagery is a clear advantage. Besides, the other output of
Facets Stereo Vision, the orthoimage, is a colour image
instead of a grey value picture.
Further tests with the modified Facets Stereo Vision have
to be carried out in order to prove if there is real image
material where the new approach offers significant
advantages. Computer-generated images where this is the
case can easily be computed. Furthermore, the possibility
