+10g
0g
with a mean of |AZ| — 4cm. This result is one indicator
that the accuracy of Facets Stereo Vision is comparable to
the accuracy which can be reached by an analytical plotter
(~ 0.09 °/oo - flight altitude), as reported earlier for close-
range imagery by [3]. Also, comparing the results numerically
to those gained from the analytical plotter as well as carrying
out a visual stereoscopic check of correctness by reprojecting
the estimated 3-D object coordinates into the digital images
with subpixel accuracy, yield similar results.
The following examples are intended to give an impression of
morphological correctness and of the consequences of terrain
noise or model deficits, respectively.
6 DISCUSSION OF OUR TEST RESULTS
In fig. 2 'Meadow' we show the reconstruction results of an
area with only low terrain noise. The shaky course of the
contours results from the very low equidistance of A — 20cm
in comparison to the standard deviations of the heights which
are in the range of az c t3cm — x: 15cm. The obtained con-
tours are another indicator that oz seems to be very realistic.
Low textured areas beside the little brook are well recorded,
the little brook itself seems to be modeled in a morphologi-
cally correct way. Fig. 2c allows a closer look at the treatment
of image disturbances. Even by using only 3 images for the
reconstruction it can be observed, that image blunders mainly
affect the residual image which belongs to the disturbed im-
age — by that way, the orthoimage is not disturbed! This
effect was already reported for parts of a reseau cross in one
image by [2].
Example 'Highway', fig. 3, shows the effect of another dis-
turbance which appears only in one image: In principle, the
effect of the driving car is not different from a blunder in
one image, except that it covers a larger area, cf. fig. 3c.
The more images take part in Facets Stereo Vision, the bet-
ter are the results. Of course, parking cars which appear at
the same position in all images, become part of the recon-
structed surface, cf. fig. 4a. The strips on the highway show
very nicely, how large grey value gradients in object space
result in low standard deviations of the estimated heights,
cf. fig. 3b. Again topographic details like the ditch and the
slope of the highway are reconstructed in a morphologically
correct way.
Contrary to the first two examples, where we treated distur-
bances in single images, the last example (fig 4) shows two in-
teresting situations where the used 22-D surface model does
not fit properly to reality. First, the building is an example for
the occurrence of discontinuities in object space, cf. fig. 4a.
Vertical parts of the surface can not be well approximated by
a 2i-D surface model - a better way could be the use of a
3-D surface model as proposed by [4]. Second, the tree in the
upper left of fig. 4a is an example for the presence of terrain
noise, typical for unleaved trees. While the shadow in nearly
isochronous exposures gives a fine texture and a good surface
reconstruction result on the ground, small height disturbances
can appear near the centre of a tree.
Depending on the aim of the reconstruction, one might want
to detect and to correct the observed gross errors. Both, 00
and cz are valuable indicators for automatic detection and
manual correction. a of the orthoimage grey values reflects
at first order the number of pixels per orthoimage grey value
facet and seems to be less important.
763
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Automatic correction algorithms should be based on a se-
mantic analysis of the observed errors. The resulting residual
images could be a helpful tool, e.g. for the calculation of lo-
cal go's or for a texture analysis to select errors caused by
unleaved vegetation.
7 CONCLUSION
We present examplary results (heights, orthoimages and qual-
ity criteria) of fully automatic surface reconstruction by
Facets Stereo Vision with a very large resolution in object
space. Simply using more than 2 images simultaneously
for surface reconstruction can avoid to meet several difficul-
ties. Further error avoidance or detection can be based upon
Facets Stereo Visions quality statements.
The results including their quality statements are a promising
starting position for further semi- or fully automatic scene
understanding tasks.
8 ACKNOWLEGEMENT
The project was supported by Deutsche Forschungsgemein-
schaft (DFG) under WR 18/1-4.
We are very grateful to Dr. Kiefer and Mr. Weiser from the
Landesamt fiir Flurneuordnung und Landentwicklung Baden-
Wiirttemberg, who made available to us the complete under-
lying data, including digital and analogue images.
REFERENCES
[1] Kaiser B. and Wrobel B. 1996. Facets Stereo Vision
(FAST Vision) Applied to digital Colour Images. XVIIIth
ISPRS Congress, Comm. Ill, WG 2, Vienna.
[2] Heipke C. 1989. An Integral Approach to Digital Image
Matching and Object Surface Reconstruction. In: Griin
A., Kahmen H. (eds): Optical 3 (three) D Measurement
Techniques. Wichmann Verlag, Karlsruhe: 347-359.
[3] Kempa M. and Schlüter M. 1993. DEM Evaluation
by an Operator and Facets Stereo Vision: A Comparison
Based on Close-Range Imagery. 2nd Conference on Optical
3-D Measurement Techniques, Zurich, Switzerland, 502—
509
[4] Schlüter M. 1994 Object Space Based Surface Re-
construction with Discontinuities — An Approach. ISPRS
Comm. Ill Symposium 'Spatial Information from Digital
Photogrammetry and Computer Vision’, vol. 30, part 3/2,
737-744, Munich, Germany
[5] Weisensee M., 1992. Algorithmen und Modelle zum
Facetten-Stereosehen. Deutsche Geodätische Kommision
C 374, München.
[6] Wrobel B. 1987. Digital Image Matching by Facets Us-
ing Object Space Models. 4th Int. Symp. on Optical and
Optoelectr. Appl. Science and Engineering, The Hague,
Netherlands, SPIE 804, 325—333.
[7] Wrobel B., Kaiser B., Hausladen J., 1992. Adaptive
Regularization Of Surface Reconstruction By Image Inver-
sion. In: Forstner W., Ruwiedel St. (eds): Robust Com-
puter Vision. Wichmann Verlag, Karlsruhe: 351-371.
