465
CIPA 2003 XIX 11 ' International Symposium, 30 September - 04 October, 2003, Antalya, Turkey
TÏ
;
ipi
tion. (a)
;ze’ from
•s. (b) 3-
» roughly
if homol-
:rarchical
ne points
quence to
; we have
principle
:s and de-
mologous
ily higher
îe rel
ie au-
large
up to
;thod.
rr and
com-
Figure 3: Interactive object segmentation for modeling.
ination. Then dense image matching is performed to yield
a dense grid of homologous point with a user-defined grid
spacing and a dense cloud of 3D points is recovered through
robust forward intersection.
3.3 Model Generation
As we need a surface model for visualization the point
cloud is triangulated to a triangular mesh. In our experi
ments we have reprojected the points to one of the images
and used 2^-dimensional triangulation algorithm, which
was readily available. For correct modeling it is however
necessary to implement a more qualified method, which is
able to produce a true 3D model, such as for example the
volumetric integration method (Curless and Levoy, 1996).
This is left for future work.
Many artefacts contain planar regions, where a dense point
mesh is not only redundant, but also inaccurate because of
the inevitable noise of the reconstruction process. At the
triangulation stage it is also possible to reduce the amount
of data and filter the noise by detecting the planar regions
in the point cloud with a robust linear regression algorithm
(an example is depicted in Figure 5). The regression algo
rithm should make use of the points’ uncertainties, which
are known from the bundle block adjustment to yield cor
rect planar regions (Schindler, 2003). The points inside of
planar regions can be removed and only the polygonal out
line triangulated. Finally the triangulated surface models
Figure 4: User interface for defining the image ordering.
Figure 5: Detection of planar regions for data reduction.
are textured with photo-texture from the original images.
The textured model of the dancing girls’ frieze (Waelkens
et al., 2000) is shown in Figure 6. Another example of a
textured model can be seen in Figure 7.
Figure 6: All 12 pieces of the ’dancing girls’ frieze recon
structed and arranged in their original positions.
4 VISUALIZATION
4.1 Managing Large Models
Automatic modeling techniques, such as for example the
one described in Section 2, deliver highly detailed 3D ob
ject models. However the resulting meshes often by far
exceed the rendering capabilities even of high-end graph
ics workstations. Even worse, if it comes to deliver those
models over the Internet, the transmission times become
prohibitive if standardized file formats, e.g. VRML97 (The
Web 3D Consortium, 1997), are used. It has therefore been
an active field of research in the last years to provide tools
to manage large and complex 3D data sets.
There are at least three concepts that are useful in dealing
with huge 3D objects:
First, the raw data (e.g., the output of some acquisition and
reconstruction process) contains much more detail than re
quired for visualization. Thus the original object can often
be replaced by a simplified one. This becomes particularly
important in large scenes, where only a small fraction of
