Fast Symmetry Detection with Deep Learning and GeConv

Abstract

n many computer vision applications, it is relevant to know the orientation of the object relative to its symmetry axis. While humans are generally very good at evaluating the reflectional symmetry and relative orientation of objects, constructing a robust symmetry detection pipeline working out of the box is still a challenging task in computer vision. This is particularly the case for irregular objects with skewed geometry and irregular shapes. In this paper, two approaches are presented for symmetry and object orientation detection based on traditional computer vision and Deep Learning. First, an algorithm is presented that allows fast computation of 2-axis reflectional symmetry of contour points and the definition of a radial convex hull, GeConv. The algorithm is tested on 1000+ image dataset from the wind tunnel and flight test experiment.