Learning functional object categories via mined event classes. 

Invited talk by Anthony Cohn

Abstract:
In this talk I will present ongoing work at Leeds, in
collaboration with Krishna Sridhar and  David Hogg on inducing
functional object categories from (qualitative) spatio-temporal
descriptions of the participating objects. First we mine event
classes by building a complete activity graph of the video,
represented as a layered graph of qualitative spatial and
temporal relations, and then induce event classes from this, in
an unsupervised manner. This makes use of a generative event
model which will be described. Having formed these event
classes, object categories can be formed by clustering those
objects which take the same roles in a particular event. We have
experimented with these techniques in two domains: a kitchen
scenario, and aircraft turnovers.