CIP A 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey
531
nts with the
the original
; projections
e appealing,
tches, which
lly only few
opacity than
performed
cStudio 5.0.
g. 4) were
whereas the
)s served to
redundancy,
include the
the object,
vhich could
:r reliability.
n obtained
exported to
it files were
:ts “ordered
the format
egl’s format
Studio. So,
iven by the
not ordered
me advan
ce: Due to
erations on
sr, and the
iduction or
io provides
the points
ised on the
ace normal,
s restricted
process is
(editing of
md editing
i editing of
one will be
lai prepro-
cessing was necessary. Since each scan was exported
completely, manual editing (deleting of extraneous points,
e.g. those on the ground around the statue) was necessary
to confine the point clouds to the area of interest. This
step reduced the number of points from about 7.4 Mio.
to about 4.0 Mio. points. The next step was to elimi
nate points that were erroneous due to distance errors
occurring along the contour (distance measured too long
because of averaging with the background due to beam
divergence). The wrong points occurred in conspicuous
structures and could be easily detected by displaying
the scans in different colors (cf. Fig. 4). Nevertheless,
this step was tiresome and time-intensive. It should be
avoided by recognizing data errors already in the raw data.
It turned out to be advisable to subdivide the whole
point cloud (or rather the object) into two separate areas,
namely the socket and the actual statue, each of them
having about 2 Mio. points. This was done because of
the following reason: The two parts show different surface
characteristics: The socket can be approximated by
planes, it has a prismatic shape, whereas the actual statue
is rather complex and its surface can be best characterized
by free-form shapes. This seemed to be important when
using the “reduce noise” operation, because the user
can choose between “prismatic shapes”, which preserves
better edges and corners, and “free-form”, which is
optimized for free-form shaped surfaces. The (horizontal)
separation-plane was chosen about 2cm above the socket’s
top surface because the noise around the latter was about
2cm.
The point cloud representing the socket was thinned
out as well as the one representing the sculpture, using a
uniform sampling of lcm-cubes: This means that space is
subdivided into equally sized cubical cells, and afterwards,
all but one point from each cell are deleted. Thus, the
sculpture’s point cloud was reduced to about 1.4 Mio.
points.
Because of their simplicity, it is appropriate to describe
the socket by a set of planes. Therefore, GeomagicStudio
provides a tool called “detect planes”, which searches the
point cloud for planes using thresholds/tolerances that
can be quantified by the user. Due to noise, this tool did
not work properly. So, the function “capture plane” was
used instead several times: It determines a set of points
“belonging” to the plane based upon a manual selection
of points, which are supposed to lie on a plane. Finally,
for each plane’s detected point set, the parameters of the
best fitting plane were calculated. The intersection of the
mathematically defined planes was done with a simple
program. The point cloud representing the sculpture had
to be filtered strongly because the noise was so high that
the calculation of the triangulation did not make sense yet
(see Fig. 5). So, the point cloud had to be filtered three
times (maximum filter setting, unfortunately this is deliv
ered as black box only). Step by step, outlying points were
eliminated and the roughness of the data was reduced,
too. As result, the noise was largely reduced (of course,
some details were lost and features were smoothed),
and a curvature-based sampling had become reason
able. Thus, the point cloud was reduced to 25% (about
320,000 points). The transition to polygon phase followed.
The point cloud was triangulated. The triangula
tion was quite satisfying at the most parts of the object’s
surface. Nevertheless, especially near the data holes,
there were gross errors in the triangulation, which had
to be repaired. For example, the lions at the front part
of the statue “grew together”. This caused the following
problem: The program supposed the three lions to be one
single surface and flipped the surface normal at one whole
lion because the surface normal got twisted between two
lions.
After coarse editing, the triangles were reduced (50%)
F
Figure 5: Triangulation of the unfiltered point cloud. Mea
surement noise leads to a rough surface representation.
from about 615,000 to about 307,000 in order to make
manual editing easier, almost without losing surface de
tails. This can be achieved, because the triangles already
represent a surface, and triangles in (approximately) the
same plane can be combined.
Then, a very time-intensive manual editing of the
triangulation followed: Holes and partial holes had to
be filled, spikes had to be removed, local smoothing was
performed, eliminating “wrong” features, local reduction
of triangles in hardly curved regions, refining the triangu
lation by more triangles in strongly curved regions, and
other operations like that. One must say that this manual
editing was necessary because of the shortcomings of the
triangulation due to using the merge of unordered point
clouds. The quality of triangulation can be expected to
be better when having a topology already in the point
phase.
After that, the number of triangles was once more
reduced to 50% and some fine edits were done. Finally,
the statue’s surface was represented by 79,634 points and
155,762 triangles.
The next step was to enter shape phase in order to
generate a NURBS representation of the object. Since
Version 5.0, GeomagicStudio provides a function called
“AutoSurface”, which automatically produces a patch
layout necessary to generate NURBS (see Fig. 6).
The user can decide how much surface detail shall be
preserved. In this case, the detail should be preserved as
good as possible.
The result of the NURBS surface was very satisfying but
the number of patches was 3,293 causing a file of about
205 MB (compared to 26 MB for triangulation). Creating
the patch layout manually would certainly result in fewer
patches. Of course, a manual patch layout might be a
little better in certain areas, but the effort would be
enormous.
Finally, it must be said that the most time of the
modelling was necessary to find out which steps had to
be done and in which order they had to be performed.
Certainly, this steps applied on this complex statue are
other ones than when modelling a relatively simple facade.
The relative high noise required much editing already at
