ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002
to the well-known Stanford bunny, typically used as a
demo object by the computer graphics community (Stan-
ford 3D Scanning Repository). The bitangent curves of
these patches were extracted and then matched based on
their invariant signatures. Fig. 14 shows the bitangent
curves for the first view. The automatically registered
bunny data is shown in fig. 15. Note that this completed
model is the result of the automated ‘crude registration’. A
fine registration based on ICP or another technique can be
used to refine it. Nevertheless, it already looks quite con-
vincing. The automatic matching (incl. bitangent curve
extraction) took about 9 min. on a Pentium III 1.1 GHz.
A second example illustrates 3D patch matching on the ba-
sis of the texture maps, i.e. on the basis of invariant neigh-
bourhoods extracted from these. It goes without saying
that for this technique to be useful the 3D acquisition de-
vice should also capture the surface texture. We have used
Eyetronics’ ShapeWare (Eyetronics www). We demon-
strate our approach for a globe. This is an example where
a shape-based approach is doomed to fail, due to the high
degree of shape symmetry. The texture with a represen-
tation of the continents and oceans breaks this symmetry
and makes it possible to automatically arrive at a complete
compilation of the object model. Fig. 16 shows two of
the 48 patches that were captured separately. As can be
seen, the overlap is rather small. Yet, more than 200 corre-
sponding invariant neighbourhoods could be found (with-
out propagation and transitivity reasoning in this case). A
detailed cutout of both patches with the matching neigh-
bourhoods is shown in fig. 17. The globe could be re-
constructed automatically based on the texture approach
alone. A view of the result is shown in fig. 18. Just
as in the case of shape-based registration, it is advisable
to apply a texture-based fine registration after this rather
crude stage. Johnson and Kang have proposed an approach
that could serve this purpose (Johnson 1999). This second
stage should then also take care of texture blending. As
can be seen in fig. 18 the original patches in our recon-
struction can still be distinguished by their differences in
texture intensities.
4 CONCLUSIONS AND FUTURE WORK
Three-dimensional reconstruction often introduces ‘wide
baseline’ problems. This can be the case at the point where
one has to find correspondences between the 2D input
views, or when one has to register partial 3D reconstruc-
tions. We have proposed solutions to both 2D and 3D wide
baseline matching problems. Ongoing work is mainly fo-
cused on issues of efficiency. A stronger integration of 2D
and 3D techniques remains to be explored.
Acknowledgements: The authors gratefully acknowledge
support from K.U.Leuven GOA project ‘VHS+’ and Eu-
ropean IST project ‘3D-MURALE’. Help by prof. Marc
Waelkens and K. Cornelis of the Kath. Univ. of Leuven
in gathering the archaeological imagery is gratefully ac-
knowledged. We also thank J. Matas for providing images.
A - 12
Figure 13: Range data from the Stanford bunny.
Fig
bun
the
for
Fig
ba:
