Abstract

   Natural scenes contain a huge  amount  of  information  if counted by
   spatial pixels and temporal frames.  However, most of the information
   is redundant because the pixels and the frames are highly correlated.
   The  optical  flow,  generated  by  motions  of  the  objects and the
   observer, contributes significantly to the  statistical regularity of
   such spatiotemporal correlations.   The  visual  system  of an animal
   such as the human is highly  adapted  to  this statistical regularity
   such that the visual sensitivity  follows  the  same  contours as the
   spatiotemporal  correlations  of   natural   scenes   in  motion,  in
   particular, along two axes: space  and  motion  instead  of space and
   time.  Furthermore, vision is  an  active  process,  during which eye
   movements interact with visual scenes and  select  images that arrive
   on the retina: pursuits and fixations on  objects significantly alter
   the image velocity  distributions  on  the  fovea  and the periphery,
   which lead to the dependence of the visual sensitivity on the retinal
   eccentricity;  saccades  between  objects  change  the natural  scene
   statistics dynamically, which lead to  the  dependence  of the visual
   sensitivity on the time relative to saccades.   All  of  these can be
   accounted for  by  the  proposed  ecological  theory  that the visual
   system maximizes the causal information of the natural visual input.

   (In: Dynamics of Visual Motion Processing (Ilg UJ and Masson GS eds),
    Page 261-282, 2010)