Symmetry similarity of human perception to computer vision operators

Abstract

Symmetry occurs everywhere around us and is key to human visual perception. Human perception can help guide the improvement of computer vision operators and this is the first paper aiming to quantify that guidance. We define the Degree of Symmetry (DoS) as the measure of ‘how symmetrical’ a region is as human perception sees symmetry in a continuous manner. A new dataset of symmetry axes, the Degree of Symmetry Axis Set, is compiled for ordering by DoS. A human perception rank order is found by crowd-sourced pairwise comparisons. The correlation of two ranked orders is defined as the Symmetry Similarity which we use to evaluate symmetry operators against human perception. No existing symmetry operator gives a value for DoS of a reflection axis. We extend three operators to give a value for the DoS of an axis: the Generalised Symmetry Transform, Loy’s interest point operator, and Griffin’s Derivative-of-Gaussian operator. The highest Symmetry Similarity of a symmetry operator to human perception revealed they are a poor approximation of human perception of symmetry.