1087
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B5. Beijing 2008
Figure 4. From top left clockwise, a) original image. The top right marble area is new with chiselled surface resulting in many
repetitive bright dots, b) the blue channel of the noise-reduced image. Noise is reduced but also some signal is filtered out. The
texture is very weak, c) Same as b) after Wallis filtering [Wallis, 1976]. d) same as c) with about 22.5 deg convergence angle. Note
the good effects of Wallis mentioned in the text, and the small repetitive texture, which sometimes differ in c) and d) due to different
viewing angle and marble properties.
For Mamiya, we tried three different parameter sets of the
Wallis filter. Visual inspection of the matching results did not
show any significant differences between the three versions.
Figure 4 shows an example of Wallis filtering. The large texture
improvement is apparent. But difficulties in matching such
images might arise. Indeed, in homogeneous areas which often
occur with marble surfaces, especially new marble, noise may
have been enhanced and appear as texture. This underlines the
importance of using cameras with good radiometric quality. The
second problem is that when the camera view changes, different
texture appears in the image due to the crystal structure of
marble and the different reflectance and automated matching
becomes impossible. This problem increases with larger
convergence angle between the images and occurs in both sun-
illuminated and shadow areas. The last problem is that the
texture is often of small size and repetitive, leading to multiple
solutions and wrong results in matching. Occlusions are an
additional problem.
4.3 Results of 3D modeling and surface generation
For the image-based 3D modeling of the Erechtheion, as the
monument contains different typical architectural elements (like
columns, flat walls, etc.), manual measurements were applied
(Figure 5). On the other hand, for some details (Karyatides,
ornaments, etc.) and to investigate image matching problems on
marble surfaces, automated approaches for dense surface
reconstructions were used. For these parts we employed a (i)
Depth from Shading (DfS) method (Figure 6) (El-Hakim, 2006)
and (ii) the ETH matcher able to recover dense and accurate
results using simultaneously multiple images (Zhang, 2005;
Remondino et al., 2008). The method combines multiple
matching primitives and various area-based matching
techniques to exploit all the content information of the images.
It can also cope to a large extent with depth discontinuities,
wide baselines, repeated patterns, occlusions and illumination
changes by using several advancements over standard stereo
matching techniques. Figures 7 and 8 show the 3D results from
the Canon and Mamiya camera respectively.
To evaluate the accuracy of the image matching results and
investigate possible penetration effects of laser light into the
marble (Godin et ah, 2001; Lichti and Harvey, 2002), we
compared the photogrammetrically derived point clouds with
the range data (Surphaser 25HSX, range accuracy specification:
lmm@15m, see www.md3d.uk.com/surphaser.html) acquired
for the range-based modeling of the entire Erechtheion (El-
Hakim et ah, 2008). The average point density of Surphaser
was 2 mm, and for the matching surface models 1 mm and 2mm
for the Canon and Mamiya respectively. The 3D comparison is
performed within PolyWorks© IMAlign and IMInspect. Firstly,
the two meshes have to be co-registered (rotation, translation
and scale, the latter estimated manually) and then a best fit
between them is computed, minimizing the distances between
the mesh patches.