Motion and disparity processing informs Bayesian 3D motion estimation

Welchman et al. (1) propose a Bayesian model that combines a velocity prior for slow motion (2, 3) with approximations of lateral velocity V_x and velocity in depth V_z to model biased perception of 3D motion trajectories in the x–z plane (3, 4). Although decomposing a motion vector into orthogonal components may be mathematically convenient it raises the question of why the visual system should approximate these velocities in the first place.

The binocular visual system has no immediate mechanism to encode retinal velocities in terms of component V_x and V_z. It must estimate both vectors through retinal motion and/or disparity input. (Note that interocular velocity difference and change of disparity are mathematically equivalent but may reflect physiologically different processing streams.) Retinal motion and/or disparity are readily available to the visual system (5), and these signals are sufficient to estimate motion trajectories in a binocular viewing geometry (3). Thus, a 3D trajectory can be estimated directly without first approximating vectors V_x and V_z and corresponding likelihoods.

Lages (3) developed Bayesian models of binocular 3D motion perception that combine a prior for slow motion in x–z with likelihood constraints of motion and/or disparity processing to explain perceptual bias of motion trajectories. By comparing three model types (motion-first, stereo-first, and stereo-motion) it was concluded that disparity rather than motion processing introduces perceptual bias for 3D trajectories ranging over 360°. These Bayesian models have the advantage that likelihoods are based on physiologically plausible motion and/or disparity processing.