Abstract
Performance in human face discrimination tasks can be degraded by manipulating the position of features (i.e. eyes, nose, mouth) within a face stimulus (Tanaka & Farah, 1993). The effect is typically attributed to a disruption of face mechanisms in the brain involving feature configuration (de Haas et al. 2016). Here, we use an ideal observer (IO) and a foveated ideal observer (FIO; Peterson & Eckstein, 2012), to investigate the extent to which that performance difference can be attributed to the interaction between the altered feature locations and the sampling of foveated processing arising from eye movement strategies. We create a fixation-weighted FIO (FW-FIO) in order to control for effects on performance caused by the observers' empirical fixation strategies. Methods: Six observers completed an emotion discrimination task with three emotions (happiness, sadness, and fear) using movies of 20 faces (15 deg. height, 1400ms presentation) in luminance noise. Four conditions were used; a control upright face, a face with inverted (upside down) features with intact locations, and two different anomalous configurations of upright facial features. Observers were free to execute eye movements. Results: The efficiency of both humans and the FW-FIO compared to the IO is significantly lower for configurations in which important features for the task (eyes and mouth) are further away from each other relative to those configurations in which they are closer together and cannot be foveated simultaneously. However, humans compared to the FW-FIO, have even lower efficiency in the altered configurations relative to the control configuration. Conclusion: Our findings suggest that much of the degradation in performance with altered facial feature configurations can be attributed to the interaction between foveated processing and the location of features relative to observers' fixations. However, there is a residual efficiency loss related to other configuration specific mechanisms higher in the visual stream.
Meeting abstract presented at VSS 2018