Seminars/Colloquia/Invited Talks

Seminars

Paolo Emilio Barbano

Non-Linear Dimensionality Reduction For the Classification of Motion-Styles

PLACE:	Clark 314
EVENT:	CIS Seminar Series
DATE:	February 7, 2006
TIME:	1:00 - 2:00

Abstract

Modern Motion Capture technology faces the extremely challenging problem of identifying characteristic parameters of "Motion Styles", i.e. identifying under which kind of conditions individuals perform the same prescribed movements.
In doing this, two substantial obstacles need to be to overcome. One is the inherent high-dimensionality of the data: essentially all motion data arises from tracking of a large number of markers labeling the limbs of the human body.
The other is the large amount of variation to be factored in while comparing data recorded from different individuals: small physical differences are source of great variance.
We propose a new Trainable Classifier architecture capable of explicitly approximating the metric of the high-dimensional manifold on which the data lives and exhibiting a global dimensionality reduction map. Examples from explicit motion styles are provided and effectiveness of the methodology is demonstrated with real motion capture data.