Purchase this article with an account.
Gary C.-W. Shyi, Julia W.-J. Lin; Effects of Exposure Frequency and Pose Variation on Learning 3-D Faces: A Comparison between Viewpoint Interpolation and Extrapolation. Journal of Vision 2014;14(10):1258. doi: https://doi.org/10.1167/14.10.1258.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
A major issue in face recognition and generalization concerns how the visual system can overcome image variations caused by pose and expression to derive stable and constant representations. In their recent study, Shyi & He (2011) demonstrated the joint influence of exposure frequency and expression variation on face learning where multiple exposures coupled with sufficient number of expression variation can yield optimal performance in face recognition and generalization. Here in two experiments we extended Shyi & Hes study and examined the effects of exposure frequency and pose variation on learning 3-D faces. In particular we compared how viewpoint interpolation and extrapolation might differ in their effects on face recognition and generalization. In Experiment 1, where each face was exposed for 12 times with different degree of pose variation during learning, we found relatively poor performance in recognition due to pose variation, and participants committed relatively high rate of false positives when recognizing faces that were never shown during learning. In Experiment 2, we found much better performance with rates of false positives substantially reduced when each face was exposed 24 times during learning. Most importantly, we found in both experiments that interpolated views not only yielded results comparable to single views when they were old, but also yielded better results when they were new, implicating robust viewpoint generalization due to interpolation. In contrast, extrapolated views, while also yielding modest level of generalization, the effect of generalization was much weaker than interpolated views. It is also interesting to note that, unlike the more dynamic nature of interaction between exposure frequency and expression variation (Shyi & He, 2011), the effects of exposure frequency and pose variation appears to be additive. Possible roles for viewpoint interpolation versus extrapolation in learning 3-D faces across pose variation are discussed.
Meeting abstract presented at VSS 2014
This PDF is available to Subscribers Only