Parameterized Models for Optical Flow
Research Overview
Techniques for estimating motion from video image sequences have
improved significantly in the last ten years. One key to the success
of many techniques is the use of parameterized models of motion that
help constrain the estimation problem. It has been common, for example,
to assume that motion between frames of an image sequence is well
described as an affine deformation. Although these models work well
with smooth surfaces and fairly simple rigid motions, there are many
other types of complex motions for which they are less effective.
We have been exploring the design and use of parameterized motion models
that represent much more varied and complex motions. Our goals are
threefold: to construct simple, linear models of complex motion phenomena;
to estimate the coefficients of these linear models directly from
spatiotemporal image derivatives; and to recognize or classify motion
``events'' from the recovered coefficients.
We are interested in three types of complex motions classes:
- Domain-specific motions: An example is the motion of human mouths
during facial expressions or speech. Motion estimation of mouths is a
challenge for conventional optical flow techniques because of the
significant changes that occur from frame to frame. Moreover, models
are difficult to hand-code, and therefore we learn them from examples
using principal component analysis.
- Motion features: Examples include motion discontinuities and moving bars
We construct a set of {\em steerable basis flow fields} that approximate the
motion features.
- Motion texture: The third class of flow that we are interested
in modelling is motion texture. Examples include the motion of leaves
in the wind, or flags fluttering. These remain extremely difficult
motions to estimate and recognize in image sequences.
Related Publications
- Fleet, D.J., Black, M.J., Yacoob, Y., and Jepson, A.D. (2000)
Design and use of linear models for image motion analysis.
International Journal of Computer Vision 36(3):171-193
(abstract)
- Fleet, D.J., Black, M.J. and Jepson, A.D. (1998) Motion Feature
Detection Using Steerable Flow Fields. IEEE Conference on Computer
Vision and Pattern Recognition, Santa Barbara, June, pp. 274-281
(pdf)
© IEEE
- Black, M.J., Yacoob, Y., Jepson, A.D., and Fleet, D.J. (1997) Learning
parameterized models of image motion. IEEE Conference on Computer Vision
and Pattern Recognition, Puerto Rico, June, pp. 561-567
(pdf)
© IEEE
Return to David Fleet's home page.