The Extraction of
Spatio-temporal Energy in Human and Machine Vision
Edward H. Adelson and James R. Bergen
Published in
Proceedings of the Workshop on Motion: Representation and Analysis (pp. 151-155)
Charleston, SC; May 7-9 (1986)
Recent work in human motion perception has conceptualized motion detection in terms of filters
selective for spatio-temporal (ST) energy. One class of such models are called
"energy models." They code motion energy, not velocity as such, but we describe a
velocity coding model based on the ST energy within given bands of spatial and
temporal frequency.
The motion energy computations appear at first to be quite different from the
gradient-based computations that have been used in machine vision. But gradient
systems that make appropriate use of confidence can be considered to derive
velocity from opponent energy measures.
We also note that in human vision, dynamic energy seems to be coded in a set of
transformed axes, which are opponent motion energy (e.g. right - left) and
counter-phase flicker energy. The flicker axis is often neglected, we note its
possible utility in visual processing.