TITLE: Image Categorization Using Local Probabilistic Descriptors

PRESENTER: Katarina Mele, http://www.cse.unsw.edu.au/db/staff/info/katarinam.html, katarinam@cse.unsw.edu.au

AFFILIATION:CSE - University of New South Wales, University of Ljubljana, http://www.cse.unsw.edu.au/, http://cogvis.fri.uni-lj.si/

DATE: Friday 16th September 2005

TIME: 12:00:00

PLACE: CSE Seminar Room K17_113

ABSTRACT:

Image categorization involves the well known difficulties with
different visual appearances of a single object, but introduces also
the problem of within-category variation. This within-category
variation makes highly distinctive local descriptors less appropriate
for categorization.

I will introduce a new type of local image descriptor called
probabilistic patch descriptor (PPD). PPDs encode the appearance of
image fragments as well as their variability within a category. We
apply PPDs to image categorization by using machine learning where
the features are the matching scores between images and PPDs.
Experimental results show the benefits of modelling the
within-category variation of local descriptors. We have experimented
with different local descriptors in combination with PPD. I will
present some results where as local descriptors RGB values, SIFT
descriptors, and the combinations of both are used.

BIOGRAPHY OF SPEAKER:

Katarina Mele is a PhD student at University of Ljubljana and a
visiting student at CSE. Katarina has spent a year with us here
at CSE to work on her thesis. Next week she is going back to Slovenia,
where she will finish her PhD. The talk is a part of her PhD thesis.

Host:

Seminar Convenor:

Van Hai Ho