667
CIP A 2003 XIX"' International Symposium, 30 September - 04 October, 2003, Antalya, Turkey
ING
itua.gr)
ith.gr)
ccuracy with
laeology still
n most cases
ified fishing-
lar strip geo-
d strong dis
ing to ensure
ion of break
being exten-
lotogramme-
nmetric mo-
5, but also to
ogrammetric
y (but some-
lvestigated.
sses. Unlike
objects may
ary’), distin-
es’, edges or
:n caused by
almost per-
ictures. Sur-
ighly crucial
ogrammetric
2002).
l is mostly a
n archaeolo-
2000). Laser
igy, capable
horter times,
:tion (Monti
the obvious
ge; noise), a
ocessing for
l orthophoto
02); perhaps
;se, the most
besides, not
;rs as it may
îd to above,
Greek sites.
a key issue
mode of 3D
reliable ex-
ial task. Re-
ndle adjust-
ment may often be infeasible, mainly due to limited extension in
object depth but also unfavourable strip geometry or inaccuracy
and low identifiability of control points (mostly simple natural
detail points). On the other hand, full pre-calibration may not be
always practicable, particularly if different non-metric cameras
are being used (as in the case of the Department of Surveying &
Photogrammetry of the Greek Ministry of Culture).
Karras & Mavromati (2001) have indicated that, in most cases,
employment of the ‘nominal’ camera parameters (ignore princi
pal point of analogue cameras; use nominal focal length as the
camera constant) does not considerably affect accuracy. But this
is not so for radial lens distortion, especially of the wide-angle
lenses. Typically, the authors pre-estimate distortion separately,
by line fitting to distorted straight linear features; this correction
has consistently yielded satisfactory results and may even treble
accuracy (Karras & Mavromati, 2001).
A first task presented here was to record a 180 m long façade of
the ancient Greek castle in Aigosthena (~300 B.C), severely da
maged by the 1981 earthquake. Because it is situated on a steep
rock surrounded by trees, horizontal recording was made using
a raised medium format Fuji camera with a 45 mm wide-angle
lens (Fig. 1). The object consisted chiefly of planar surfaces, de
velopments of which were finally mosaicked. Yet a certain part
showed considerable depth, thus requiring orthorectification (cf.
Fig. 5). Photogrammetric 3D modeling was inevitable, as under
the circumstances it was not possible to employ a laser scanner.
Based on 6 images (mean scale 1:300), the mean RMS error of
bundle adjustment in XYZ for the 68 control points was 1.3 cm
(2.1 cm without correction of distortion).
Figure 1. The meteorological balloon at the Aigosthena Castle.
The results are indeed satisfactory (radial distortion of the parti
cular lens was rather small). Other aspects of successful adjust
ment are outlined in Mavromati et al. (2002), notably problems
due to poor adherence to flight planning resulting in demanding
image geometry. As shown in Fig. 2, in both vertical (Karras et
al., 1999) and horizontal photography rotations about the verti
cal axis are the least controllable. Image recording on a mildly
windy day in February produced large (J)-tilts up to 15° (Fig. 3).
Figure 2. Critical rotations (a: vertical, b: horizontal images).
In such situations, rather small stereo-bases are needed to secure
adequate overlap, along with ample control and tie points, mea
sured carefully on the image with its strong perspective distort
ions due to surface relief and image tilt. Differences in imaging
distances are also to be kept within certain tolerances. Besides,
dense recordings are necessary also for avoiding occlusions and
ensuring the required photo-texture; in fact, here all six images
have been actually used in the process of orthorectification.
Figure 3. Footprints of the six horizontal images.
3. GENERATION OF 3D SURFACE MODELS
Locally imprecise description of complex surfaces causes ‘erod
ing’, ‘stretching’ and ‘melting’ effects on the resampled images.
Most commercial software, as the one used here, describe object
surface as a 2.5D DSM, i.e. with a single elevation value at each
planimetrie location (fully 3D description requires special soft
ware; e.g. Knyaz & Zheltov, 2000). Thus, photogrammetrically
collected heights and breaklines are typically integrated by a 2D
Delaunay triangulation into a surface mesh. In our experience, a
most usual problem in archaeological orthoimaging is modeling
surfaces almost orthogonal to each other (formation of ‘vertical’
triangles). In such cases, the software must be ‘assisted’ by suit
able collection.
As a protection against ‘arbitrary’ triangulation, Mavromati et
al. (2002) have reported on a collection scheme depicted in Fig.
4. For each segment d of a breakline, representing the top edge,
three points are collected at the bottom: two correspond to its
endpoints (A, C) and one (B) is close to its middle. Though un
questionably tedious, this process ‘forces’ triangle formation to
adapt itself faithfully to surface form.
Figure 4. Breakline and points forming ‘vertical’ triangles.
Figure 5. Initial image and partial view of the shaded model.
In Fig. 5 the initial image shows the morphology of the object;
also, a partial view ot'the 3D model is also seen. It is clearly ob
served that surface edges have been faithfully modeled to ensure
geometrically correct orthoprojection. Examples of orthoimages
