Hierarchical
object recognition (discussion topic)
C. Williams
Some older AI ideas wrt object recognition consider schemes
where one
first groups low-level features into subcomponents, and these
subcomponents are then grouped into components recursively, up to
objects. Currently applied methods for object recognition seem to use
very shallow hierarchy, as in e.g. the constellation model, or in
templates for SIFT features (e.g. Sudderth et al, ICCV 2005; Fergus et al,
ICCV 2005). It might be interesting to have a discussion session as to
if we should be aiming for more hierarchy. Some recent examples
of models including more hierarchy are Bouchard and Triggs (CVPR 2005),
and the "composition machine" (S Geman, CVPR 2006).