Abstract
The face-space framework models face representations as locations in a multidimensional space endowed with a dissimilarity-based metric. We recently demonstrated that similarity (distance) measurements in this space under one viewpoint corresponded to similarities under another viewpoint. These data propose that tolerance of identity representations across viewing conditions can be interpreted as "tolerance of similarities". To link this notion to behavioral phenomena of view-tolerant face processing, a behavioral and a computational experiment were carried out. In the behavioral experiment, subjects rated similarities separately within each of two variants of a set of faces, differing in viewpoint either by 90o, 60o, 30o or 0o. In each condition, we constructed two face-space configurations, one for each viewpoint, and compared them. We found that these configurations were similar across different viewpoints. Interestingly, correspondence was significantly lower between configurations differing by 90o in viewpoint than between other configurations. Next, subjects performed a "same"/"different" matching task across viewpoints. Performance mirrored the different degrees of correspondence across configurations observed earlier. We further hypothesized that this tolerance of similarities could causally underlie the observed matching performance across viewpoints, if every face image was encoded as a vector of similarities to its viewpoint-specific prototypes. To test this hypothesis, we generated faces under the previously employed viewpoints, and a computational model evaluated their similarities to arrays of randomly chosen prototypes within each view. We then simulated a matching task, with images having highly correlated patterns of similarities to viewpoint-specific prototypes judged to be the "same" face. Results indicated that the scheme was feasible and appealing, with performance decreasing with increasing magnitude of viewpoint change. Hence, tolerance may be established via within-viewpoint similarity judgments, instead of through direct comparisons across viewing conditions, which has been standardly advocated by both human and machine vision models.
Meeting abstract presented at VSS 2012