Vision and Robotics Seminar
Amnon Shashua
School of Computer Science \& Engineering
Hebrew University
will speak on
Learning over instance spaces made out of sets
with application to video sequence interpretation
Abstract:
The talk is about developing a similarity function that operates on pairs of
sets of vectors --- where a vector can represent an image and a set of vectors
could represent a video sequence for example --- in such a way that the
function can be plugged into a variety of existing classification engines. The
crucial ingredients are therefore (i) the function can be evaluated in high
dimensional spaces using simple functions (kernel functions) evaluated on pairs
of vectors in the original (relatively low-dimensional) space, and (ii) the
function describes an inner-product space, i.e., is a positive definite
kernel.
We will be using two running examples to motivate and test the ideas described
in the talk. The first task is the discrimination of ``suspicious'' motion of
an individual or a group of individuals in a video sequence. We use the SVM
approach using our similarity function defined over matrices where an input
matrix represents the motion trajectory of a group of individuals over a
certain (fixed) time frame. We show that the classification (suspicious versus
non-suspicious) greatly outperforms the conventional representation of forming
a single vector out of all the trajectories. The second application is the
visual recognition of f aces from input video sequences representing head
motion and facial expressions.
This work is co-authored with Lior Wolf.
The lecture will take place in the
Lecture Hall, Room 1, Ziskind Building
on Thursday, November 14, 2002
at 11:00