| Abstract: |
Image contours make excellent features for the modeling and
recognition of shape-based object classes (e.g. horses, or
mugs). Recently, LEAR has developed a technique to learn the common
boundaries shared by multiple examples of an object class (e.g. the
mugs' handle), while discarding contours arising from the details of
the individual examples (e.g. logos drawn on the mugs). Currently,
this learning technique requires the training examples to be marked by
bounding-boxes. The goal of this DEA is to develop a novel method
which can learn simply given the images, without giving the location
of the example objects. This would substantially reduce the level of
supervision needed by the algorithm, and hence ease the task of
collecting training data.
|