[4eyes] FW: 2/23/17 Vision Seminar Sven Dickinson, University of Toronto

[Matthew Turk]

In the midst of all the faculty candidate talks, there will be an interesting computational vision talk by our frequent lab visitor Sven Dickinson (University of Toronto) next Thursday - 1pm in 3834 Psychology East. I encourage Ilab members to attend.

	Matthew

Vision Seminar - Thursday 23rd -   3834 Psychology East 1-2:30pm.

The Perceptual Advantage of Symmetry for Scene Perception

Sven Dickinson
Department of Computer Science
University of Toronto

As one of the original Gestalt principles, symmetry is believed to support visual perception by aiding the visual system in detecting objects, which tend to be symmetric. Whereas the role of symmetry for the perception of isolated objects has been well studied, it is so far unknown what role symmetry plays in the perception of cluttered, real-world scenes. We demonstrate, for the first time, a strong perceptual advantage of local contour symmetry for perceiving complex real-world scenes. Unlike global symmetry, local symmetry is largely invariant to pose. Scenes were represented as line drawings, which have been shown to capture the essential structural information required for successful scene categorization (Walther et al., 2011). We assessed local symmetry by computing the degree to which pixels participate in a non-accidental symmetry relation of the scene, the medial axis transform (Blum, 1973; Siddiqi et al, 2008). Each contour section was assigned a numerical symmetry value based on the rate of change of the radius function of the medial branch to which it was assigned. We then generated two alternate versions of each line drawing, one with the half of the pixels ranked most symmetric and one with the half ranked least symmetric. The two types of modified line drawings were shown to twelve participants along with intact line drawings in a six-alternative forced-choice scene categorization experiment with short presentations (53 ms), followed by a perceptual mask. Each participant saw 20 images from each category per condition (360 total trials). Participants’ categorization accuracy was significantly higher for the most symmetric contours (49.7%) than for the least symmetric contours (38.2%), with intact line drawings showing higher performance than both modified conditions (65.8%). These results demonstrate, for the first time, the role of local contour symmetry as a crucial organizing principle in complex real-world scenes.