Item

Abstract

In examining 3D motion estimation (1), we start by considering the computation of two entirely plausible quantities—changes in disparity (δ) and azimuth (φ). Because these depend on the difference and mean of retinal signals, respectively, we expect φ to be more reliable. The impact of this differential reliability on motion estimation is best illustrated for pure motion in depth (Vz) and pure lateral motion (Vx). Vz depends only on the higher-variance δ signal, whereas Vx depends almost solely on φ for small movements. Thus we measure performance along these dimensions to illustrate the limitations imposed by the underlying computations.

Lages and Heron (2) query whether it is biologically plausible for the brain to estimate Vx and Vz separately. As we have shown elsewhere (3), formulating the estimation problem in terms Vx and Vz does not imply that the brain estimates the components separately (although performance can clearly be measured along these dimensions as we have shown). Rather, trajectory angles could be calculated directly based on δ and φ (4). The Bayesian model we use (in ref. ,1, see figure 1C) selects the estimator from the two-dimensional (Vx, Vz) space rather than relying on independent estimation.

Lages's previous work (5) suggested that a model based on static disparity (δ) provides the best account for 3D motion estimation. Although this is a formulation we have used previously (,3), we suggest that, within the context of judging moving objects, it is more plausible that estimation be based on motion rather than a disparity snapshot.