ul 2004
blished
elected
les.
tion
a least
58) that
paring
nber of
ich the
tk. The
pipolar
les and
rates a
ametric
nsional
its are
igorous
Figure 9: Inclusion of edge information
The final 3D point cloud prompts the update of the Delaunay
triangles. Upon examining the updated 3D surface model,
potential model fidelity discrepancies due to discontinuities and
gradient changes can be successfully addressed by introducing
edges. The single pixel width edges in this example have been
extracted using a canny edge detector and converted into a
linked list of edge points so that they can be processed in a
similar manner to the initial interest points. To achieve this edge
detection is initialised in a seed image and the subsequent
process of creating 3D edge segments proceeds in the manner
described for the interest points. Once computed in 3D, the
points are grouped to produce 3D edge segments which may
then be used to constrain the existing triangulated surface. This
method provides a high level of model fidelity.
3.1.1 Summary of the multi-photo surface extraction method
e A multi-photo convergent image set are used to
automatically measure the Retable surface and
generate a surface model
e Network adjustment ensures correct image geometry
e A generate basic triangulated surface model from
target point data is produced (such a model is
consistent in both image and object space)
e Seed interest points are identified (Foerstner operator)
on a triangle by triangle basis
e Multi-photo epipolar constraints are used in
conjunction with triangle boundary information to
identify search regions for possible homologues in
other images
* Multi-photo patch matching with geometric
constraints ensure accurate homologous point
correspondence
e An intersection solution at a specified tolerance level
validate the data and produce 3D coordinates
e Surface triangles are automatically updated at each
iteration
3.2 Surface measurement summary
The described approach to surface densification provides a
dense surface point cloud with sub-millimetre level precision.
This level of precision is supported with results of the network
adjustment and has been validated against known engineering
surfaces in the laboratory (Papadaki 2002). Quality of the
derived model data is further supported by the fidelity of the
model when inspected in 3D with the calibration corrected
imagery draped onto the triangle network.
The technique can resolve issues such as occlusion by
integrating measurements from different views. Fine detail can
be successfully included by careful selection of minimum image
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
AIME FIGE
processing patch sizes and careful image acquisition at an
appropriate magnification. Spurious data, due in the case of the
Retable, to repetitive decorative patterns are filtered out by the
geometric and radiometric constraints employed throughout the
process.
Ultimately, the method is limited by the image network
geometry and the existence of and appropriate surface views
and texture content within the images. In this case, the use of
edge detection has provided significant advantages over other
methods due to the nature of the surface of the Retable, which is
characterised by sharp gradient changes and linear features
produced by paint layering and wood frame designs.
Figure 10: Rendered surface model of the Retable
4. THE ART CONSERVATION DATABASE
Photogrammetric surveys were undertaken primarily to assess
the mechanical response of a complex object to its environment.
However, as a result of these surveys, a 3D model of the
Retable was created that offered other functionalities. These
additional functions — static measurements and spatial
referencing — are a consequence of the existence of a
dimensionally accurate digital image.
A spatially referenced image of the Retable enabled the creation
of a database that facilitated the documentation of the
conservators work. Traditionally, a conservator will generate a
written record of their work as it progresses. This is document is
supplemented by the results of technical analysis and visual
records such as X-rays, Infra Red reflectographs, and
photographic images of the object before, during and after
treatment. If the object and the treatment are relatively
straightforward, involving relatively few conservators working
over a period of months, then such methods of documentation
are completely appropriate. However, the complexity of the
Retable and its treatment required project management support
software that did not exist. Using principles related to those
employed in GIS software, a conservation database was built.
A hierarchical description of the Retable was created as a
framework for the five thousand or so separate components of
which it is comprised. This included all original components
(from the six oak planks to several hundred small gems) and all
later additions (the wooden reinforcements attached to the back
at the turn of the nineteenth century, etc). The boundary of each
individual component was also outlined on the digital image (or
multiple images — one for the front, one for the reverse, one X-
ray mosaic, etc).
The conservators could then approach the Retable as an
assembly of individual spatially related components. Access
was either via a standard directory-structure type hierarchy or
