Notes from Meeting on July 7, 2003
Linda's explanation of the project
Paralleling Sal's work with the bunnies and clustering of 3D regions
that were specified by hand, we are going to use the regions from the
color, structure, and texture files to cluster together sets of
regions. We will classify each region and presumably use both overlap
and known semantic characteristics to then correctly label each
region. For the clustering, we will probably use k-means and maybe
the EM (expectation maximization) algorithm.
Sal's idea
Sal thinks that this clustering will not be as clean as we might hope,
and is anticipating that by introducing the idea of semantic
detectors, such as window dectectors, to recognize various classes of
objects. Detectors that we would use would need to be robust against variance in :
- scale
- brightness
- perspective
.
Linda has pointed out that scale might not matter--as long as line
segments are visible, Yi's structural output should have detected and
outputted the sides of any structure that contains segments. Also,
Yi's structural algorithms already seem to handle perspective by using
the vanishing point of certain segments and then trying to make them
parallel, or something along those lines.
Consensus of the approach for the group
For now, we are going to focus on continuing with Yi's work. His
approach has not yet been fully explored and the way he is trying to
go about clustering and classifying objects is one that no one else is
taking right now. Sal's idea can be worked on in parallel, but is not
going to be the focus of the group until we have at least run a
preliminary round of experiments on what Yi's processing yields.
Immediate tasks
Our next goal is to run a test on a subset of the images collected by
Clifford and Jenny (as well as using the images that we already had in
the database) to determine how well the structural output is able to
distinguish the presence and absense of structure. The thought is to
use 500 images with structure and 500 without, then calculating the
number of false positives, false negatives, and accurate readings, so
that we can determine the confidence that we can have in the
structural output. Clearly, we need to keep in mind while running
this test that the density of structure detected, as well as color and
textural features, will also be taken into consideration when we are
doing the final classification of each type of object.