9.357 Topics in Vision Science: Natural Scene Statistics
Outline: Texture Segmentation, posted 10-01-03

Download (same text as below)

This week (9/30) I talked predominantly about statistically-based models of
pre-attentive texture segmentation (Julesz, Beck, etc.) Next week (10/7),
I'll start off talking more about filter-based models of texture
segmentation. As you might recall, two complaints with statistical models
were that they weren't actually that specific about what statistics might
be used, and we didn't know how the brain or a computer might implement
them, given images as input. Otherwise they can be kind of cool, because
they can give you good intuitions about what textures will and will not
segment. Filter-based models are sort of the opposite. We know how to
implement them, but they don't give us very good intuitions about what will
segment and what won't.

In an attempt to both have a better understanding of what statistics might
be involved in texture segmentation, and to bring together these two
classes of models, I'll talk about some work I did on this problem a few
years ago. First, psychophysics experiments aimed at getting at what
statistics determine segmentation of "orientation-defined" textures. Then,
how to implement that statistical model using the same sort of filterbank +
non-linearities model as previous filter-based models.

-Ruth ( r r u t h <a> M_I_T_e_d_u )

Classes
9.357