A model of figure-ground segregation from kinetic occlusion

Learning to Find Object Boundaries Using Motion Cues

“…Traditionally studied in controlled laboratory experiments [4,8,26,42], such motion cues have received comparatively little attention for the purpose of object popout in computer vision, e.g. [3,7,29,34].…”

Occlusion boundary detection and figure/ground assignment from optical flow

“…A particle filter approach is employed to deal with the complex, multi-modal distributions in the highdimensional state space. [8] uses local motion information only to infer boundaries and direction-of-figure. In [11], tensor voting yields optical flow estimates together with an uncertainty measure based on the homogeneity of the votes.…”

“…Occlusion boundaries are assumed to be maxima in the uncertainty measure. Neither [4], [8] nor [11] make use of static cues. Since the optical flow as a secondary feature requires integration over a spatial domain to deal with the aperture problem, edges based only on motion estimates are usually inaccurate and dislocated.…”

Motion feature detection using steerable flow fields

“…Another approach uses filters tuned to the local spatiotemporal image structure of a motion discontinuity [6,8,11]. These methods model image structure over time rather than the motion field.…”

“…s (x;c) t ; 1) ; I(x t ) ) : (6) Here, is a scale parameter and ( ) is a robust error function applied to the residual error r(x;c) = I(x +ũ s (x;c) t ; 1) ; I(x t ) : (7) For the experiments below, (r ) = r…”