Abstract
The brain often combines sensory measurements in a fashion approaching statistical optimality. If sensory measurements are represented as likelihood distributions, the optimal combined percept lies in between two single-cue percepts. When these measurement distributions differ greatly from one another (cue conflict), it may be more sensible to simply use the more reliable (less variable) measurement rather than a combination affected by the less reliable measurement, even if it is not statistically optimal. We investigated how disparity and texture are combined in estimating slant, and in particular how large conflicts affect slant percepts. Previous depth cue combination work has been restricted to small conflicts. We also investigated the effect of cue reliability.
We measured two-cue slant percepts using a 2-interval matching procedure. One interval had a cue conflict and the other did not. We presented a large range of conflict sizes. As conflict size increased, observers' slant settings remained nearly optimal: up to surprisingly large conflicts, they followed the weighted average predicted from single-cue measurements and Bayes' Law. With even larger conflicts, the optimal weighted averaging no longer occurred; matches were instead dictated by the more reliable cue. When we increased the reliability of both cues, keeping the relative weights fixed, optimal weighted averaging broke down at smaller conflicts, compared to the less reliable stimuli.
Thus, we found that optimal weighted averaging occurs when these distributions overlap substantially: either with smaller conflicts or less reliable cues. Slant percepts are dictated by the more reliable cue when the distributions only overlap minimally.