การติWe have related the disparity of residual pixels to two
geometric constraints, namely the epipolar constraint and
structure consistency constraint. The disparity is used to
filter out the parallax regions from the initial detection results.
Unlike other proposed parallax filtering methods that
directly threshold the disparity, we express the disparity
values in probabilistic terms, namely likelihood functions.
Then the filtering result is formulated as a likelihood map
instead of a binary mask image.
The likelihood models of a pixel to belong to an independently
moving object based on the geometric constraints are
defined as follows: