In computer vision research some work is done to separate 
specular reflection components from Lambertian components 
(e. g. Coleman and Jain (1982) and Klinker, Shafer and 
Kanade (1988)). A few attempts have been made to directly 
extract parameters of either one or both of the reflectance 
models (e. g. Nayar, Ikeuchi and Kanade (1988) and Ikeuchi 
and Sato (1990). The experiments provided in these investi- 
gations are usually based on simulations. 
Stereo orthophotos 
The terms stereo orthophotos or stereo orthoimages we want 
to use in a distinct analogy to the term stereo images. From 
the two images taken in standard stereo configuration the 
two corresponding orthophotos are derived by the same pro- 
cedure. The overlapping area in the images yields redun- 
dant orthophoto information pixel by pixel. These over- 
lapping areas of both orthophotos are what we call the 
stereo orthophotos. To avoid confusion it is notable to say, 
that this is not in agreement with the understanding of the 
term stereo orthophotos as it is used in some textbooks 
of photogrammetry!. Originally the idea behind stereo or- 
thophotos was to simplify the mapping process. For this a 
second copy of an orthophoto was generated, in which the 
image points are shifted with parallaxes reflecting the eleva- 
tion model. 
For updating existing maps Peterle (1989) uses orthophotos, 
which are derived from images taken of the same object but 
with a time lag of several years. This implies, that the scene 
in meantime in general has changed. Even though the map- 
ping is of prior interest in this work, there are similarities to 
the verification task as we will see later. 
Viewing the stereo orthophotos stereoscopically, as proposed 
by Finsterwalder (1985) mainly for orthophoto verification, 
can be an efficient way to verify the DTM manually. Just in 
the case of working at digital photogrammetric workstations 
this seems to be more promising than other techniques like 
superimposition of wireframe or contour line representations 
of the DTM on original stereo imagery. Up to now superim- 
position is realized only as an analog technique, applied in 
the environment of analytical plotters, which work with the 
photographs directly. 
Verification 
Without doubt the human endowment with the ability to 
see, to discern and locate objects, to reconstruct and under- 
stand the 3D space is the splendid pre-condition to do the 
verification of DTMs using stereo orthophotos. The geomet- 
ric model for parallaxes between the orthophotos is just the 
expectation that the parallaxe at any point is zero. This is 
equivalent with the expectation, that the spatial impression 
gained from the stereo orthophotos is a planar object which 
corresponds to the plane of the orthoprojection. All spatial 
deviations from planarity, i. e. all nonzero parallaxes, in- 
dicate inconsistencies between the D'TM and the real world 
surface. 
With this paper we want to pick up the complex of auto- 
matic DTM verification. The way to develop a fully auto- 
matic D'TM verification procedure is presumably quite long, 
  
!Sometimes the term stereo orthophotos is used to characterize two 
orthophotos which are derived from one image by orthoprojection along 
different spatial directions of projection, or which has the same effect, 
by shifting all image points of an orthophoto by parallaxes proportional 
to the elevation difference of the point over a reference plane (Blachut, 
1971, Collins, 1968). 
234 
because at least some of the human abilities of understanding 
images have to be incorporated in the procedure. A strategy 
to approach the verification task could be as follows: 
(1) Detect intensity differences between the orthophotos. 
Supposed the orthophotos are generated in such a way that 
intensity values are obtained at identical location, the corre- 
spondence between the orthophoto pixels is implicitly given. 
In a simplest subtraction method, the intensities of each pixel 
of the stereo orthophotos are subtracted from each other. 
Significant nonzero values in the difference orthoimage indi- 
cate inconsistencies between the orthophotos. By aggrega- 
tion of the affected pixels into areas, directly the location of 
doubtful areas is indicated. If we assume that the physical 
aspects of image formation addressed above are not the rea- 
son for these changes, then the differences can be interpreted 
as indication for discrepancies between real world geometry 
and DTM. More sophisticated methods than the simple sub- 
traction are discussed in the next section. 
(2) The areas of interest located in the first step may be hints 
for objects like trees, houses, or bridges which are not repre- 
sented in the DTM. Furthermore structural information like 
breaklines or other discontinuities, which are not included in 
the measured data or which are smoothed out within surface 
interpolation, also might be the reason for the differences. 
Such errors and all other geometric errors which have not 
been recognized within the data capture or the interpolation 
process are candidates for areas of discrepancies. The in- 
terpretation involves modelling, reconstruction and location 
estimation of the 3D structure or shape of the objects or im- 
perfections of the DTM. The next step is to decide if and if 
necessary which of the reconstructed information has to be 
included or eliminated from the existing DTM. In this clas- 
sification constraining rules (e. g. do not represent mobile 
objects, trees, etc.) have to be incorporated. It is obvious 
that other informations like colour would be helpful if not 
even necessary to come to an automation of this interpreta- 
tion task. 
In this paper we present some work on the localization of 
conflicting regions between two orthophotos. The term lo- 
calization as used in this context comprises the detection of 
discrepancies and, in consequence, the location of regions in 
the DTM which do not agree with the real world surfaces. 
The idea is to analyse the whole area covered by stereo or- 
thophotos and locate the discrepant regions. These areas are 
marked and visualized, so that instead of a detailed visual 
control of the whole area we only have to carry out inter- 
pretation for a certain percentage point (e. g. 5 % or 1 %) 
of the total area. In the next section we briefly describe 
three different procedures for localizing discrepancies in the 
orthophotos and in the third section we will present results 
with simulated and real stereo imagery. 
2 LOCALIZATION OF 
DISCREPANT REGIONS 
BETWEEN STEREO 
ORTHOPHOTOS 
All procedures we describe work on the iconic level of im- 
age description, i. e. the intensities of the orthophotos are 
used for comparison. In consequence, the problem we have 
to solv 
betwee 
in the 
age for 
inated, 
tested. 
cated c 
closing 
crepan! 
DTM i 
intensi 
terior c 
do is t« 
individ 
sequen: 
are fulf 
thopho 
positio: 
functio! 
But for 
approxi 
The pr 
ancies ¢ 
the ster 
the ort} 
respecti 
radiome 
The adj 
only gla 
This m« 
dure sig 
For moc 
the ster 
(FE) me 
The adj 
is true. 
C(m, n) 
ing of t] 
facets wi 
contribu 
lation o] 
individu 
joint edg 
with mo 
face C is 
to a cert 
models kh 
context 
from dig 
For the t 
mation h 
ent from