|
| |
        |
|
From a practical
point of view, it would be impossible to predict the appearance of every
target feature in the scene due to noise, occlusion, and segmentation
errors. The solution is to predict the appearance of
groups of features only. This is straightforward using AFoVs:
simply combine reference views containing the corresponding model groups.
The key question is how to choose the model and image groups. We need to
select image groups containing features from the same object only, a
challenging problem especially when the scene contains clutter. |
| |
|
To deal with this
problem, we propose integrating AFoVs with grouping. Grouping can be
defined as the process of clustering geometric features into higher level
descriptions that are likely to have resulted from the same object. The
key idea is to consider object properties that are very unlikely to occur
at random. Among the properties that have been proposed (e.g., colinearity,
parallelism, proximity etc.), convexity is an important one. It has been
shown that convex groups of edges rarely occur at random and are
very likely to have resulted from the same object. Moreover, convexity is
invariant to perspective projection. Using convex grouping for recognition
assumes that the targets have convex shapes or contain convex parts which
is not very restrictive. |
| |
|
