Dirk Bernhardt-Walther, Psychology Department, St. George Campus, University of Toronto
Title: Which Features Do People Use to Categorize Natural Scenes?
Ebbinghaus Empire Series 2014-2015
VIDEO RECORDING OF TALK
Abstract: People can categorize their natural environment as, say, a forest, a highway, or an office, very quickly and accurately. What image properties underlie this remarkable ability? We extracted a number of candidate features from line drawings of natural scenes and measured their relevance for this task. Computational analysis showed that the statistical distribution of contour orientation is the most informative feature. Comparison of the structure of categorization errors between computational and behavioral scene categorization, however, revealed that contour junctions and curvature but not contour length or orientation are related to human performance in categorizing scenes. When we manipulated the line drawings such that contour junctions got perturbed, however, participants resorted to using contour length and orientation. Comparisons of error patterns between the computational analysis and decoding of scene categories from fMRI activity showed that the parahippocampal place area (PPA) primarily relies on junction properties and curvature in its representation of scene categories, whereas primary visual cortex is mostly related to contour length and orientation. We conclude that people are able to opportunistically use a range of features for scene categorization, but given the choice, they will use contour junctions and curvature. These features also underlie the representation of natural scene categories in the PPA. Importantly, these features are well suited to construct structural descriptions of shape rather than global statistics of image texture.
For further information please contact Morgan Barense at email@example.com.