The problem of finding groups in a set of data is one of the main problem of
pattern analysis.
This is also important in vision, since it is well known from the results
of the Gestalt school, that the main perceptual law are grouping laws.
When making some hypotheses on the data (number or shapes of the groups for
instance), efficient clustering methods exist. The problem of rejecting the
existence of any group (when necessary) is more seldom tackled in clustering
analysis.
In this talk, I will present the so-called a contrario detection framework,
following which a meaningful group is defined as a large deviation of an
hypothesis of independence on the data. Two general problems are of
particular interest: first, how to decide that a group candidate is
"valid"? Second, in the case of nested groups, how can we choose the best
representative ones?
This will be illustrated in the one-dimensional case (histograms).
I will then present of generalization to higher dimensions, applied to the
problem of finding groups of similar pieces of shapes in two arbitrary
images.
Presentation (PDF File)
Video of Talk (RealPlayer File)
Copyright. All Rights Reserved.