face but to focus on the goal of surface geometries and
on those methods that are best suited for a particular
need. As a result one may find today commercial
systems for modeling smaller industrial objects or
human faces that rely on non-stereoscopic ranging or
even tactile reconstruction systems. An industry not
much smaller than that of photogrammetric equipment
has recently emerged that will feed the geometry of
objects into computer graphic rendering systems.
These may be based on magnetic, acoustic or optical
tracking of a cursor in 3-D space. It may employ
ranging, or use structured light. Table 1 is an attempt
at summarizing the range of modeling techniques for
3-dimensional objects.
Stereoscopy
optical and active echo-ranging, electron-microscopes
Tactile profiling
magnetic tracking
optical tracking
acoustic tracking
Structured light
Exploiting geometric constraints in single image
- Ranging with lasers and radar altimetry
Interferometry
Shape-from-Shading
Use of shadows, layover
Photometric stereo
Tomographic imaging
Table 1: Techniques Used for Extracting
Object Surface Geometry
1.3 Surface Radiometry
The extension of reconstruction from a purely geo-
metric to a global view of surface properties is a result
of the transition to so-called „Digital Visual Informa-
tion“ to encompass both ,Image Processing” and
„Computer Graphics“. The surface consists of the bald
reference object, objects placed on the bald reference
surface, information about the surface’s reflective
properties such as color, specularity and texture.
Methods of assessing the surface radiometry are well-
understood in remote sensing. The brightness infor-
mation obtained in an image needs to be inverted to a
measure of reflective properties of the surface. This
benefits from multiple looks at each surface point in
different spectral channels as well as from different
vantage points.
The assessment of texture is most commonly accom-
plished from photographs. The object geometry is
being modeled by polygons and each polygon re-
ceives a photographic facet. This facet needs to be
made independent of effects of the illumination and
viewing direction valid at the time of imaging.
1.4 Rendering a Surface
Clearly surface reconstruction is a function of its
application. Traditionally photogrammetry has been
used for topographic mapping so that the resulting
maps are available for navigation, orientation and
planning. With the advent of computer generated
images and computer graphics the surfaces also are
being used for rendering. The problem is then ex-
panded by issues of illumination, viewing position
and viewing direction. The map as traditional product
is being replaced by geometric and radiometric object
422
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
models and by the graphics pipeline, to feed the
softcopy visualization of an object.
2 STEREOSCOPIC RECONSTRUCTION
2.1 A Systems View
Stereoscopy is the best understood method of surface
reconstruction in photogrammetry. The traditional
view decomposes stereoscopic machine vision into a
set of individual worksteps (Table 2).
Image orientation
Image matching
Preliminary surface definition
Match verification and acceptance
Final surface representation
Gridding
Data formatting
Table 2: The stereoscopic process flow.
The most widely discussed aspect of stereoscopic
machine vision is image matching. The least under-
stood is verification and acceptance of a surface point.
2.2 Image Matching
The search for homologue features in overlapping
images has received most of the attention spent in the
past on automated digital stereoscopy. A recent
authoritative presentation is by Forstner (1993). The
matching domain is being addressed essentially in
image or object space and in terms of feature matching
or area-based matching. The last 25 years have seen a
proliferation of techniques that focus on the speed of
matches, the robustness and independence of radio-
metric and geometric disparities in images, the pull-in
range, i.e. the ability to find a match even if the
geometric differences between two images are large,
the smart prediction of presumed match locations, the
idea of using resolution pyramids in a hierarchical
approach, the optimization of similarity measures, e.g.
in terms of a least squares estimation of the match
location, the combination of geometric and radio-
metric parameters to determine a match, the reduction
of the dimensionality of the problem by constraining
the search areas along epipolar lines.
The matching ideas further could be grouped into
those applicable when nothing is known about the
object and the camera (Zhang et al., 1995); when metric
cameras are used and the orientation of the cameras in
3-D space is known; when the object is fully 3-dimen-
sional and has many hidden and occluded elements;
when structures exist that can be approximated by
polygons, e.g. when looking for buildings in an urban
environment; when the radiometry interferes with geo-
metry such as in active echo-ranging systems (radar
and sound).
The matching accuracy is variously reported as
ranging between + 2 pixels in highly dissimilar image
pairs such as those obtained with a speckle-infested
radar system to + 0.05 pixels or better when sharply
defined, rotationally symmetric objects can serve as
homologue features in overlapping, well-illuminated
stereo photography using retro-targets.
The
SCO]
veri
min
tati
cee
suc
bei
ima
Yet
WO!
wil
for
3.1
The
tai
par
zen
din
urb
str
not
to
ph
knc
roc
wh
ide
fro
cep
cer
ext
gra
3.2
In
ave
tio:
ave
foc
me
prc
gra
the
im
2-c
de
the
(516)
3.3
Of
noi
jec
als
nai
cla
sec
pr