Abstract
We've known for over a century that estimates of pursuit speed are typically lower than estimates of retinal speed. The benchmark is the Aubert-Fleischl phenomenon, the name given to the perceived slowing of moving objects when they are pursued. Many other pursuit-related phenomena can be accounted for in a similar way: stationary objects appear to move (the Filehne illusion), motion trajectories are misperceived, perceived heading oscillates (the slalom illusion) and slant increases. When compared quantitatively, the slowing required to explain these phenomena is remarkably consistent. Estimates of pursuit speed are evidently lower than estimates of retinal speed – but why? Recent Bayesian accounts of retinal motion processing may provide an answer. These rely on a zero-motion prior that reduces estimates of speed for less reliable motion signals. This would explain the phenomena above if signals underlying eye-velocity estimates were less precise, a prediction tested by comparing speed discrimination for pursuit (P) and fixation (F). Using a standard 2AFC task, we found trials containing P-P intervals were harder to discriminate than F-F intervals. We also found that speed matches for F-P intervals revealed a strong perceived slowing of pursued stimuli (Aubert-Fleischl phenomenon). A control experiment showed that poorer P-P discrimination was not due to the absence of relative motion. We used a Bayesian observer to fit psychometric functions to the entire data-set. The model consisted of a measurement stage (single non-linear speed transducer + two sources of noise), followed by a Bayes estimator (SD of prior free to vary). The model fit the data well. In order to explain the other phenomena listed above, however, the model demonstrates that estimates of retinal motion and pursuit must be added after the Bayes estimation stage. Adding signals beforehand (at the measurement stage) cannot predict changes in velocity, just speed.