Abstract
The ability to detect faces in images has high ecological significance. It is a pre-requisite for important tasks such as person identification, gender and affect analysis. However, systematic experimental studies for characterizing the limits of human face detection performance have so far not been undertaken. We address this problem through a combination of techniques from psychophysics and computational vision. Our data provide lower bounds on image resolution needed for discriminating face from non-face patterns and help characterize the nature of facial representations used by the visual system. Specifically, our experiments allow us to derive the following inferences:
-
The lower bounds on image resolution needed for a particular level of face-detection performance: Faces can be reliably distinguished from non-faces even at just 2 cycles eye to eye using only the internal facial information. We have also analyzed how performance varies across different kinds of non-face distractors.
-
The role of local context in face-detection: The inclusion of facial bounding contours substantially improves face detection performance, suggesting that the internal facial representations encode this information.
-
The role of luminance contrast polarity: Contrast polarity appears to be encoded in the representation since polarity reversals have significant detrimental effects on detection performance, particularly with inner features. The visual system is more tolerant to contrast negation in the presence of bounding contours perhaps by encoding these contours in a contrast invariant manner.
-
The role of image orientation: Changes in image orientation away from the upright decrease face detection performance. Our data additionally suggest that vertical bilateral symmetry per se may not be a significant determinant of face detection performance.
Current efforts focus on computationally modeling this body of experimental data.