Robust visual estimation as source separation

Abstract

We developed a method analogous to classification images that allowed us to measure the influence that each dot in a dot cluster had on observers' estimates of the center of the cluster. In Experiment 1, we investigated whether observers employ a robust estimator when estimating the centers of dot clusters that were drawn from a single distribution. Most observers' fitted influences did not differ significantly from that predicted by a center-of-gravity (COG) estimator. Such an estimator is not robust. In Experiments 2 and 3, we considered an alternative approach to the problem of robust estimation, based on source separation, that makes use of the visual system's ability to segment visual data. The observers' task was to estimate the center of one distribution when viewing complex dot clusters that were drawn from a mixture of two distributions. We compared human performance to that of an ideal observer that separated the cluster into two sources through a maximum likelihood algorithm and based its estimates of location using the dots it assigned to just one of the two sources. The results suggest that robust methods employed by the visual system are closely tied to mechanisms of perceptual segmentation.