Abstract
Humans can optimally integrate sensory cues across different perceptual modalities in order to form a coherent percept (Ernst & Banks, 2002). It is unclear, however, to what degree optimal integration also occurs within a single perceptual modality. Here, we argue that humans integrate visual motion signals across independent spatiotemporal frequency bands and combine them with an general expectation for slow speeds (Stocker & Simoncelli, 2006). We formalize this hypothesis with a Bayesian observer model containing multiple sensory channels, each sensing motion energy within a limited spatial frequency band. When triggered by a moving stimulus, the responses of individual channels undergo divisive normalization and are then integrated across all channels tuned to the same speed. We tested this model against data from a range of psychophysical experiments using broadband sinusoidal gratings with different frequency spectra. Fits of the observer model to individual subjects' data well account for the full set of psychometric functions, thus can accurately predict perceptual biases as well as discrimination thresholds. Based on these results, we propose that optimal cue combination is a plausible mechanism for the integration of independent sensory signals into a coherent percept—even within a single perceptual modality such as visual motion.