Bias in visual motion processes: A theory predicting illusions

The first step in the interpretation of video sequences is theestimation of spatio-temporal image representations, that is
the image motion or the correspondence of points or lines.
Noise in the data, however, causes a serious problem for
their estimation; in particular, it causes bias. As a result the
location of image points, the orientation of image lines and
the length and orientation of image motion vectors are esti-
mated erroneously. It is suggested that machine vision sys-
tems processing video should include statistical procedures
to alleviate the bias. However, complete correction in gen-
eral does not appear to be feasible, and this provides an ex-
planation for most well-known geometrical optical illusions,
such as the café wall, the Zöllner, the Poggendorff illusion
and other recently discovered illusions of movement.