The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
no difference between a view rendered from the model and a
photograph taken from the same viewpoint, is generally
required and obtained with the texture mapping phase. This is
generally referred to as appearance modeling. Photo-realism
goes much further than simply projecting a static image over
the 3D geometry. Due to variations in lighting, surface
specularity and camera settings, colour and intensity of an area
shown in images taken from separate positions will not match.
Measurement of surface reflection properties (BRDF) and
illumination photometric measurements should also be included
for better texture modeling. The images are exposed with
whatever illumination existed at imaging time. This
illumination may need to be replaced by illumination consistent
with the rendering point of view and the reflectance properties
(BRDF) of the object. Also the range of intensity, or dynamic
range, in the scene can sometimes not be captured in a single
exposure by current digital cameras. This causes loss of details
in the dark areas and/or saturation in the bright areas, if both
coexist in the scene. It is thus important to acquire high
dynamic range (HDR) images to recover all scene details
(Reinhard et al., 2005), e.g. by taking multiple images using
different exposure times.
3.4 Visualisation of the 3D results
The ability to easily interact with a huge 3D model is a
continuing and increasing problem. Indeed model sizes are
increasing at faster rate than computer hardware advances and
this limits the possibilities of interactive and real-time
visualization of the 3D results. The rendering algorithm should
be capable of delivering images at real-time frame rates, at least
20 frames-per-second, even at full resolution for both geometry
and texture. For large models, a LOD approach should be used
to maintain seamless continuity between adjacent frames.
Luebke et al. (2002) and Dietrich et al. (2007) give a good
overview of this problem.
4. IMAGE-BASED MODELING OF THE
ERECHTHEION
4.1 Image acquisition and orientation
The image data were acquired with two SRL digital cameras: (i)
a Canon 5D (12 MPixel) equipped with a 24 mm lens and 8.2
microns sensor pixel size; (ii) a Mamiya ZD Digital Back (22
MPixel) equipped with 45 mm lens and 9 microns sensor pixel
size. Mamiya was used only for some test sites with a ground
sampling distance (GSD) of 0.5 - 0.8 mm. The Canon was used
for imaging and modeling the majority of the monument with a
quite varying GSD. Both cameras were pre-calibrated in the lab,
using the software iWitness (www.photometrix.com.au). The
calibration of Mamiya was quite older than the time of image
acquisition and thus not so accurate. The Mamiya images were
employed mainly for modelling of the whole Acropolis from a
balloon (will not be covered here) and in this work for research
purposes, to test the potential of large format (48x36 mm) CCD
cameras and to compare the image matching results from the
two cameras and with the results from laser scanning. Thus, for
Mamiya 6 test sites were selected and at each site, 5 images
were taken, of which only the central 3 were used for matching
(frontal and two convergent with an angle of about 22.5 deg).
For the Canon images, only a few signalized and geodetically
measured control points existed, due to limitations on targeting
on a historic monument and difficulty to access the highest
parts of the monument. These points were used to georeference
the final 3D model. For the Mamiya images, only a signalised
scale existed (see Figure 8) and the surface models from these
images were transformed to the model from laser scanning via
the procedure described in Section 4.3. For practical reasons
(mainly for manual modeling and use of few images), most of
the images were acquired with wide baseline and relatively
large convergent angles. This resulted in significant occlusions
and light variations between the images.
4.2 Image quality and pre-processing
Figure 3. Top: original image with significant
pattern noise visible in homogeneous areas and
unnatural colour. Bottom: pre-processed image with
noise reduction. Also the colour saturation was
reduced and the brightness increased to generate
more “natural” images (this however does not have
an influence on matching). The images show a part
of the used scale bar.
Typically, before applying the ETH matcher (Section 4.3) a
Wallis filter (Wallis, 1976) is used to enhance the texture,
brighten-up the shadow regions and radiometrically equalize
the images, making matching easier. If the images are noisy,
before Wallis, the noise is first reduced. While this was not
necessary for Canon, Mamiya had significant pattern noise. In
modem cameras, often sharpening functions are applied to
make the image visually more appealing, this however increases
the noise and introduces edge artefacts, both negative for
automated image-based measurements. Thus, for Mamiya, first
a strong noise smoothing was applied, but only to the lightness
component not for the colour. Figure 3 shows a part of the
original and pre-processed images. After this processing the R,
G, B channels were inspected visually, and their histogram
statistics examined. It came out that the blue channel had less
noise, slightly better contrast and better definition of edges,
with R being the worst channel. Thus, the B channel was
selected for further pre-processing and matching.