ON MULTI-IMAGE RECONSTRUCTION FROM HISTORIC PHOTOGRAPHS
I. Kalisperakis b , M. Rova a , E. Petsa 3 , G. E. Karras b
a Department of Surveying, Technological Educational Institute of Athens, GR-12210 Athens, Greece (petsa@teiath.gr)
a Department of Surveying, National Technical University of Athens, GR-15780 Athens, Greece (ilias_k, gkarras@central.ntua.gr
KEY WORDS: Calibration, Orientation, Non-Metric, Bundle, Reconstruction, Visualization, Architectural Heritage Conservation
ABSTRACT
The ultimate purpose of this work is to reconstruct photogrammetrically a distinguished building in the centre of Athens, which has
been tom down years ago. Five old photographs were available taken apparently with the same camera, the nominal calibration para
meters of which could be somehow ‘guessed’. As a first step, it was decided to employ our own bundle-adjustment software, which
works within a commercial CAD environment. Besides, unlike most commercially available 3D reconstruction software, it allows full
control over the whole adjustment process (by presenting individual image point residuals, producing RMS errors for check points,
accommodating additional calibration parameters etc.). A second task was to compare these results with PhotoModeler, in order to
evaluate this widely used (but in certain aspects ‘obscure’) tool against a rigorous photogrammetric approach. Besides solutions with
no control information, a few full and partial control points were established from existing architectural drawings, allowing self-cali
bration procedures. The two software tools produced essentially equivalent results, thus validating the precision of the PhotoModeler
approach. However, certain additional features of a proper bundle adjustment program (e.g. recovery of radial lens distortion, self-ca
libration with minimal or unconventional control), discussed here, may allow a fuller exploitation of the powerful reconstruction and
visualisation tools of the PhotoModeler type. It was confirmed that, using suitable software, rigorous approaches can be applied to
historic images, and results of reasonable precision may be expected, limited only by possible inaccuracies in scale.
1. INTRODUCTION
Working with historic images (and regrettably this is not all that
rare) represents one of the most challenging aspects of architec
tural photogramme try. The facts that these mostly depict objects
which no longer exist and, hence, cannot be surveyed or photo
graphed anew; that very often no control information is at hand;
that the cameras may be unknown; that these photographs have
mostly been taken at random - these facts point to the difficulty
facing the photogrammetrist. Single-image techniques, to which
C1PA Task Group 2 is dedicated, is a topic most studied as re
gards typically structured man-made objects (Brauer-Burchardt
& Voss, 2001; Petsa et al., 2001; van den Heuvel, 2001). To a
certain extent, one might say, the methods could be somewhat
standardised in such cases. If overlapping images are available,
more options are open, depending on the existing additional in
formation. Double-image or multi-image solutions may be con
sidered, either employing rigorous photogrammetric approaches
or using the commercially available (essentially user-oriented)
3D reconstruction tools.
At the present example, both options were taken using the well-
established PhotoModeler (3.1) tool and our own bundle adjust
ment software. Commercial programs, like 3DBuilder or Photo
Modeler, have two main aspects. On the one hand, they provide
tools allowing fast production of results. Besides, although it is
highly questionable whether non-expert users can in fact handle
even mildly complicated cases, such programs have greatly con
tributed to the acceptance of photogrammetry in architectural or
archaeological documentation. On the other hand, however, as
they do not address the photogrammetrist, they are in no need to
offer a full documentation of the algorithms. As a consequence,
certain aspects remain obscure, and at some points one may not
be sure as to how results have been obtained, which is the actual
accuracy, whether any assumptions have been made etc. In the
framework of the 3D reconstruction of a demolished building in
Athens, several of the above issues have also been investigated.
Built in 1870 by an unknown architect, the ‘Tsopotos’ residence
was, until its demolition, a characteristic Athenian house, partly
imitating a particular ancient Greek monument (Fig. 1).
Figure 1. The ‘Tsopotos’ residence painted by the distinguished
poet and painter N. Egonopoulos (Egonopoulos, 2001).
2. IMAGES AND SOFTWARE
2.1 The historic photographs
Five medium format historic photographs of this building were
available (Fig. 2). Although it is was obvious that they had been
acquired at different times, the photographer was known to the
source (EA1A Archives), and it was assumed that in all probabi
lity the same camera, with a normal 80 mm lens, had been used.
This was confirmed by using two vanishing points on images,
which (ignoring the principal point) are sufficient for an estima
tion of the camera constant.
2.2 The bundle adjustment program
Besides PhotoModeler, as mentioned already, our own bundle
adjustment software was used. This particular program, named