PERCEPTIONSeminarsMachine learning methods and imprecise probabilities for computer vision
Thursday May 4
, 11h00
- 12h00
, room F107
Machine learning methods and imprecise probabilities for computer vision
Computer vision is an intriguing field in which difficult problems need to be attacked using a manifold of
advanced mathematical tools.
In this talk we will discuss how some of the most interesting vision problems (action and gesture
recognition, identity recognition, emotion classification, object tracking, data association) can be
approached by means of a number of formalisms developed in fields as diverse as machine learning, Riemannian
geometry, stochastic modeling, and non-additive probabilities. We will briefly describe, among the others,
how to use bilinear models of hidden Markov models for invariant motion classification, and propose a
learning scheme in which the "best" Riemannian metric for a training set of dynamical systems is used to
classify image sequences. On the other side, we will introduce the theory of non-additive probabilities and
illustrate the way they can be naturally employed to fuse different image features or a-priori information
to increase the robustness of tracking algorithms. As one can expect, the interplay of theory and
applications turns out to be pregnant with new results on both sides.
