CIPA 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey
2.4 3D detail scanning with a GOM ATOS II
Because the marble reliefs show very fine details, it was neces
sary to use a high precision scanner for their documentation. A
GOM ATOS 11 scannner was chosen. This scanner projects
fringe patterns onto the object and uses two cameras to analyze
the resulting images (GOM 2003). Since high resolution was
important, the version with a 400 mm base and 35 mm cameras
was selected. In this configuration, the scanner yields about 1.3
million points in a field of view of 175 mm x 140 mm. Thus,
twelve scans would cover one relief (not counting numerous
additional scans which were needed to reduce the hidden areas
due to occlusions). The raw data for one single relief amounted
to about 450 - 700 Mbytes.
2.5 Photogrammetric imaging
A photogrammetric documentation of the whole object was
carried out by a private surveying company experienced in the
documentation of cultural heritage. A Zeiss UMK metric
camera was used. In addition, stereo images were acquired for
each relief on high resolution b/w film. Also, orthogonal images
were exposed on color film for later rectification and/or
texturing.
3. HARD- AND SOFTWARE FOR PROCESSING
3.1 Hardware
Since very large amounts of data have to be handled and
visualized, a 2.667 GHz PC with 1.5 Gbyte RAM, including a
GeForce4 4600 video board with 128 MB RAM, was acquired
for data processing.
3.2 Software
As reported earlier (Marbs, 2002), software requirements for
large 3D models consisting of irregular meshed surfaces are
very demanding. Some products will not even load more than 1
million 3D points let alone meshing and editing of such data.
From all programs tested for this task, Raindrop Geomagic
Studio (which was used in the latest Versions 4.1 and 5) proved
to be the most versatile (Geomagic, 2003). Nevertheless, even
with this software it is not possible to run all processing steps
(see following section) when a complete model for one of the
82 cm x 55 cm reliefs is loaded.
Thus, for the time being, the following proceeding had to be
chosen for the 3D representation of the whole cenotaph:
• one coarse model with 2 mm sampling for the whole object
using Mensi S25 data
• one tine model based on GOM ATOS data with 0.3 mm
basic sampling (further reduced by about 40% using a
curvature based algorithm) for every one of the 24 reliefs
• five to six very fine partial models for every one of the 24
reliefs with 0.1 mm sampling using the full resolution
available from the GOM ATOS II data
4. DATA PROCESSING
4.1 High resolution model (one model per relief)
Merging. In a first step, all (up to 35) scans available for one
relief are imported into Geomagic Raindrop Studio. After all
points outside the relief area are deleted, the various scans are
merged into one single data set. This does not include any
transformations since the registration was already completed
earlier in the GOM software using the targets which were
determined in the photogrammetric densification process
described in section 2.2.
Thinning. Point density in the object varies considerably. Areas
in the foreground may be registered in many scans taken from
different viewing angles whereas areas in the background may
have been scanned only once (or even been missed completely).
The aim of the thinning procedure is to delete surplus points in
repeatedly scanned areas. At the same time, the total number of
points has to be reduced below 4 million points which is a
critical value for some of the following procedures, especially
the reunion procedure following the hole filling (‘Merge
Polygon Objects’, see below). After some experiments, the best
solutions could be achieved when the points were first thinned
to a uniform sample width of 0.3 mm and subsequently the
sampling rate of a curvature based algorithm was changed until
the required threshold of 4 million points was reached. This
could be accomplished by deleting about 35 to 45 % of the 0.3
mm sample.
Meshing. The automatic meshing procedure creates about 8
million triangles from of the 4 million points. With the hard-
and software used, this will take about 10 to 15 minutes of
processing time.
Checking manifold meshes. At the end of modeling, the object
should be covered by one continuous mesh only. After the first
meshing, several isolated meshed surfaces can result, however.
For example, the surface of a shield held by a hidden knight’s
arm may result in such a separate surface. If the surface is
significant (as in the case of the shield) it has to be connected
manually to the main mesh by introducing suitable triangles.
The result should be checked again for manifold surfaces.
Accordingly, the examination for manifold meshes should be
repeated after any manipulation of the data as described in the
following steps.
Cleaning. The cleaning procedure of Geomagic Raindrop re
adjusts neighboring triangles which show large orientation
differences. It applies a sophisticated shape-cleaning algorithm
that alters the triangulation of the point data and results in a
certain extent of relaxation of the mesh. The cleaning procedure
can handle 6 million triangles only. Therefore the relief has to
be split in two parts which are processed separately.
Hole filling. Even with many scans from different angles, some
parts of the reliefs remain unrecorded. This is due to the very
detailed structure of the reliefs which contain very sharp edges
and even free standing figures and objects whose rear sides
cannot be inspected from any observation point. The recording
method of the scanner requires any surface area to be visible
from the light projector as well as from the two cameras at the
end of the instrument’s base. This fact results in additional
inevitable holes in the object.
Before the holes are filled, the relief is divided into six parts
because the filling procedure may prove problematic if more
than 2 million triangles are loaded.
Geomagic Raindrop offers an automatic hole filling procedure
which interpolates new points based on the curvature of the
surrounding area. This works well when the area is flat and
curvature is changing smoothly. Often the last points recorded
at the edge of a hole show large deviations, however, because
they have already been partially occluded by the object parts
