Purchase this article with an account.
John Wilder, Morteza Rezanejad, Kaleem Siddiqi, Sven Dickinson, Allan Jepson, Dirk Walther; Measuring local symmetry in real-world scenes. Journal of Vision 2018;18(10):749. doi: 10.1167/18.10.749.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Symmetry is an important principle for grouping visual information in complex scenes. Last VSS, we (Wilder et al., 2017) showed that symmetry greatly influences human observers' ability to categorize a scene. We designed a method for measuring "ribbon symmetry" in a line drawing and showed that observers categorized scenes much more accurately when presented with half-images containing the most symmetric pixels compared to those containing the least symmetric pixels. Ribbon symmetry focuses mostly on parallelism in a scene. In the real world, parallel lines tend to converge due to linear perspective. To avoid penalization of this tapering of parallel lines, we here present modifications to our previous measure of symmetry in order to test to which type of symmetry the visual system is most sensitive. Our original method looked at how the radius of a maximal inscribed disc changes relative to neighboring maximal discs, and compared this change to a threshold. The number of discs in a local region that exceeded the threshold became the symmetry score. We now use the derivative of the radius function along the symmetric axis between two contours as a continuous method that does not require setting a threshold. While largely matching the results of the original method, this new method allows us to capture tapering contours by considering the second derivative of the radius function as well. Now, 3D parallel lines that are not parallel in 2D, such as those receding to the vanishing point in a highway scene, also receive a strong symmetry score. This broader definition of symmetry captures scene properties relevant for human scene categorization.
Meeting abstract presented at VSS 2018
This PDF is available to Subscribers Only