Purchase this article with an account.
Benjamin Balas; Using innate visual biases to guide face learning in natural scenes: A computational investigation. Journal of Vision 2009;9(8):477. doi: 10.1167/9.8.477.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Newborn infants preferentially orient to and track human faces. While multiple studies have confirmed the existence of an early preference for natural and schematic face stimuli, there is as yet no clear agreement as to the underlying mechanism supporting this preference. In particular, two competing theories (known as the “structural” and “sensory” hypotheses) conjecture fundamentally different biasing mechanisms to explain this behavior. The structural hypothesis suggests that a crude representation of face-specific geometry (three dots specifying eye and mouth locations) is responsible for the exhibited preference. By contrast, the sensory hypothesis suggests that face preference may result from the interaction of several generic visual preferences for qualities like dark/light vertical asymmetry (“top-heaviness”), midline symmetry, and high contrast. Despite a wide range of experimental manipulations designed to tease apart these two conjectures, a key issue regarding these proposals has yet to be addressed: Can a robust “face concept” be learned using mechanisms like these? In particular, are they useful for finding faces in cluttered natural scenes?
In the current study, I describe a computational investigation of how well different biasing mechanisms support face learning given naturalistic inputs. Specifically, I adopt a computer vision approach to assessing the utility of multiple visual biases for finding faces in complex, natural scenes. Each candidate mechanism was applied to a large database of images containing faces and performance was measured in terms of the proportion of “face fragments” successfully retrieved by the mechanism in question. The results demonstrate that the intersection of multiple generic visual constraints (with no explicit face template) is more selective for faces than the classic 3-dot pattern of 1st-order face geometry. Moreover, the properties of the face fragments recovered by each mechanism suggest novel lines of inquiry regarding the nature of infants' “face concept.”
This PDF is available to Subscribers Only