ISPRS Commission III, Vol.34, Part 3A ,,Photogrammetric Computer Vision“, Graz, 2002 
  
m. 
  
Figure 4: The features that could be matched in each of the 
3 views of fig. 3. This intersection of the pairwise matching 
sets is quite small: only 16 features remain. 
transformation mapping A; to A» and get a first approx- 
imation of the real B5. This approximation can then be 
refined by maximising the similarity between B and the 
deformed By. We call this process region propagation. If 
By is not close to A,, or not on the same physical surface, 
a good similarity is unlikely to arise between the generated 
region and By, so this case can be detected and the propa- 
gated region rejected. The propagation approach strongly 
increases the probability that a feature will be matched be- 
tween a pair of views, as it suffices that at least one feature 
in its neighborhood is correctly matched. As a result, also 
the probability of finding matches among all images of a 
set increases. 
The second idea to obtain good quality multiview feature 
correspondences is to exploit redundant sets of matches be- 
tween view pairs, or put differently, the transitivity prop- 
erty of valid matches. In our 3 view example, instead of 
only matching between the view pairs (1,3) and (1,2), 
we can also match 2 to 3. This introduces precious, addi- 
tional information. For example, if a feature gets matched 
in (1,3) but not in (1,2), we can look if it is matched in 
(2, 3). If it is, at least one of these conclusions is wrong. 
Following a majority vote, we can conclude that the lack of 
a match in (1, 2) was a failure and obtain a correct feature 
correspondence along the three views. 
In summary, starting from pairwise matches, many more 
can be generated. Of course, the validity of propagated 
and implied matches is an issue, and one has to be careful 
not to introduce erroneous information. More elaborated 
schemes to achieve this are the subject of a forthcoming 
paper, which currently is under preparation. The strategies 
proposed here are akin to recent work by Schaffalitzky and 
Zisserman (Schaffalitzky 2002). In contrast to their work, 
there is less emphasis on computational efficiency. In par- 
ticular, adding transitivity reasoning to the propagation of 
matches renders our approach slower, but it also adds to 
the performance. The combined effect of propagation and 
transitivity reasoning for our example is illustrated in fig. 5. 
The number of matches along the three views has more 
than tripled. 
2.3 Approach for dense correspondence search 
The matching of invariant neighbourhoods is only the first 
step in the search for correspondences. Good 3D models 
require the selection of dense, pixelwise correspondences. 
In the shape-from-video pipeline, the initial, sparse cor- 
ner matches provide epipolar constraints, that simplify the 
subsequent dense correspondence search. Within this wide 
baseline setting, it are the invariant neighbourhoods which 
provide the epipolar constraints. But also with these con- 
straints in place, dense correspondence search under wide 
baseline conditions requires adaptations. Although our 
current dense correspondence algorithm (Van Meerbergen 
2002), which is based on a kind of dynamic path search 
along epipolar lines, performs quite well under changes 
that are a bit larger than the ones between subsequent video