September 2015
Volume 15, Issue 12
Vision Sciences Society Annual Meeting Abstract  |   September 2015
Searching through the hierarchy: Modeling categorical search using class-consistent features
Author Affiliations
  • Justin Maxfield
    Department of Psychology, Stony Brook University
  • Chen-Ping Yu
    Department of Computer Science, Stony Brook University
  • Zelinsky Gregory
    Department of Psychology, Stony Brook University Department of Computer Science, Stony Brook University
Journal of Vision September 2015, Vol.15, 9. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Justin Maxfield, Chen-Ping Yu, Zelinsky Gregory; Searching through the hierarchy: Modeling categorical search using class-consistent features. Journal of Vision 2015;15(12):9. doi:

      Download citation file:

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

  • Supplements

This study describes how the hierarchical level at which a target is specified affects categorical search—the search for any member of an object category. Participants (n=24) searched for a text-cued target at either superordinate, basic, or subordinate levels (counterbalanced over participants). Stimuli were images of objects from ImageNet. Consistent with previous work, search was guided most strongly to categorical targets cued at the subordinate-level, but basic-level targets were verified the fastest. Our computational method first built bag-of-words histograms using SIFT and hue features extracted from 100 exemplars for each of 68 categories. To describe categorical guidance, we then quantified the feature variability within each category by finding the mean pairwise chi-squared distance between all 100 histograms. Doing this for each category, we found the highest intra-class feature variability at the subordinate level, followed by the basic and superordinate levels, the same pattern found in the behavioral time-to-target-fixation data. To describe categorization, we averaged the 100 histograms for each category, then used k-means to find those “class-consistent” features in the averaged histogram having the highest means and the lowest variability. We found the greatest number of class-consistent features at the basic level, replicating the advantage observed in verification times. Basic-level categories may therefore be verified faster because they have more consistent features to aid in the decision of category membership. Our work advances this story by quantifying these principles using features extracted directly from images. Using these same features, this work also offers an intuitively appealing explanation for differences in categorical guidance; the less variability in a target category the stronger the guidance. This simple framework shows that target guidance and verification are two behaviors that can be derived from the same features, thereby further bridging the narrowing gap between search and categorization.

Meeting abstract presented at VSS 2015


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.