August 2014
Volume 14, Issue 10
Vision Sciences Society Annual Meeting Abstract  |   August 2014
Puffball Part Segmentation: Psychophysical and Statistical Evaluation
Author Affiliations
  • Nathaniel Twarog
    Massachusetts Institute of Technology
  • Edward Adelson
    Massachusetts Institute of Technology
Journal of Vision August 2014, Vol.14, 893. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Nathaniel Twarog, Edward Adelson; Puffball Part Segmentation: Psychophysical and Statistical Evaluation. Journal of Vision 2014;14(10):893.

      Download citation file:

      © ARVO (1962-2015); The Authors (2016-present)

  • Supplements

The segmentation of silhouettes into parts is a basic, long-standing problem in the comprehension, representation, and analysis of shape. Various approaches have been proposed which utilize two-dimensional shape properties; given that both the stimulus and output are two dimensional, use of two-dimensional analyses makes intuitive sense. The success of such approaches, however, has been limited. We propose a different approach, which models the selection of silhouette parts by first mapping the silhouette to a canonical three-dimensional shape, and then performing a simple and intuitive analysis of this new three-dimensional analogue. As previously described, Puffball inflation provides us with a robust, intuitive, scale invariant three-dimensional shape (Twarog et al. 2012), and a simplified application of the 3D Minima Rule (Hoffman and Richards, 1984), identifies part boundaries that can be projected onto the original silhouette. Using a dataset of human segmentations on 80 real-world silhouettes collected by DeWinter and Wagemans (2006), we present several psychophysical and statistical evaluations comparing the part-line predictions of Puffball to those of two approaches based on minima of negative contour curvature, the Necks and Limbs algorithm (Siddiqi and Kimia, 1995) and the Short Cut Rule (Singh et al., 1999). In each of these evaluations, Puffball part segmentation performs as well as or better than the minima-based approaches; most notably, Puffball identifies part-line endpoints more consistent with human judgments that the competing models. These results suggest that the two-dimensional patterns of part-line segmentations on which existing Minima Rule-based approaches are based, though valid as observations, do not necessarily comprise the computational chain through which such part-line judgments are made. As Puffball part segmentation exhibits many of these same patterns, we propose that these observed two-dimensional patterns are in fact the indirect effects of a three-dimensional computation.

Meeting abstract presented at VSS 2014


This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.