Monday October 27th 2003, 4:00pm
Ernst and Banks (2002 Nature 415 429 - 433) provided evidence that the brain integrates haptic and stereo cues in a statistically optimal way--the weights given to each cue when presented together were proportional to the reliabilities obtained for each cue alone. Other recent experiments have also supported this result. However, there are everyday examples that seem at odds with this view. For example, our ability to see depthful scenes depicted on television or cinema screens, or monitors used in psychophysics experiments, suggests that sometimes very reliable cues, such as stereo, can be ignored or vetoed. The visual system therefore also appears to be statistically robust in the sense that it can ignore a reliable cue if it conflicts with other cues. This is true even if these other cues are less reliable but are perhaps given greater weight because together they are in close agreement. The maximum likelihood estimation method used by Ernst and Banks does not provide a good framework for modelling such types of cue integration. We present an alternative model that is both robust and optimal. This model is similar in principle to the robust ideal-observer model (Porrill et al, 1999 Nature 397 63 - 66). We describe an experiment designed to test the integration of conflicting and consistent stereo and texture cues from stereograms of a set of parabolic ridges shown at a range of orientations. The dependent variable was perceived ridge amplitude. We show how this model can account for the data obtained if blur cues to the flatness of the monitor are incorporated. The model also provides a natural way of accounting for individual differences within the data set. (Buckley D. & Frisby J.P. (2003), ECVP 2003 abstracts)
David Buckley
from the Department of Psychology
University of Sheffield
David R. Simmons