A Bayesian ideal observer for perceiving heading and rotation from optic flow

Abstract

Retinal optic flow produced by self-motion is a function of both the linear translation of the observer and the rotation of viewpoint due to eye and head movements. A challenge in analyzing optic flow is identifying the translational and rotational components. Previous psychophysical work has found that extra-retinal input from eye movements helps the human visual system extract observer translation direction (heading) from optic flow with a rotational component. However, heading can also be accurately perceived in the absence of extra-retinal feedback when optic flow is sufficiently rich. We developed a Bayesian ideal observer model to account for such results. Given an input velocity field, the model computes the likelihoods of different combinations of heading and rotation, assuming a rigid environment and noise in velocity measurements. The likelihood function from optic flow is combined with a likelihood function representing extra-retinal information about eye movements, and the maximum of the resulting function is interpreted as the observer's perceived heading and rotation. With plausible noise assumptions, the model can simulate human heading perception across a range of conditions, including: simulated vs actual eye rotations, environments with various depth structures, and the presence of independently moving objects.