Abstract
We used an equivalent noise paradigm to examine how the human visual system pools local motion signals across space to estimate global direction, and to determine what stimulus attributes limit that process. Specifically, we had observers estimate the overall direction (clockwise or counter-clockwise of vertical) of a field of moving band-pass elements, whose directions were drawn from a wrapped normal distribution. By estimating the smallest discriminable change in mean-direction as a function of directional variability, we were able to infer both the precision of observers' representation of local direction (i.e. additive internal noise) as well as their efficiency at combining local-motions (i.e. multiplicative internal noise). We estimated equivalent noise for various numbers of moving elements occupying regions of various sizes. We report that internal noise is determined wholly by the number of features present in the display, irrespective of their spatial arrangement. Crucially however, performance deteriorates faster than equivalent noise predictions at high levels of directional variability. This breakdown in observer performance can be explained by supposing that direction integration is achieved by “second-stage” channels that pool motion energy across a limited range of directions (a process that has been notionally linked to the operation of neurons in cortical area MT); overall direction is then determined by the identity of the most active channel. By incorporating elements of winner-take-all and vector-averaging approaches this model is able to account for data from all seven conditions using a single channel-bandwidth/multiplicative noise setting. We conclude that direction integration employs channels that combine densely-spaced input from local motion detectors across a wide area of space but a limited range of directions. These channels are limited by multiplicative noise and, within this context, direction integration is a strictly within-channel process.