D-Dynamics - Discriminative Dynamic Model
Learning
Traditionally, dynamic models are learned to best represent
the “observations” such as the silhouettes on an object
moving in the sequence of video frames. The “true” object
state, it’s position, pose, velocity, is missing and not
available to the learning algorithm.
However, the advancement of measuring methods in recent
years has brought changes to this traditional setting. For
instance, motion capture tools allow us to supplement the
video observations of a person performing an action (e.g.,
walking) with the “true” estimates of her pose. It now
becomes possible to use both the observations and
the targets to learn those dynamic models. Yet, the
modeling and algorithmic methods that would let one do so
are still in their infancy.
In this work we explore the space of Discriminative Dynamic
Models, dynamic models that are specifically learned to
make accurate predictions of an objects state (pose,
position, velocity,...) from the image and video
measurements. We focus on efficient and scalable learning
methods that makes use of, possibly small, sources of
labeled dynamic data. To do so we draw an analogy between
this family of models whose states are multivariate
real-valued vectors and the now famous discrete state
models such as HMMs and CRFs.