Purchase this article with an account.
Yunfeng Li, Zygmunt Pizlo; Establishing 3D symmetry correspondence in asymmetrical perspective images. Journal of Vision 2013;13(9):446. doi: 10.1167/13.9.446.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Symmetry correspondence is analogous to stereo and motion correspondence problems. Unlike the other two, symmetry correspondence was completely ignored in prior research. The importance of this problem is related to the fact that 3D symmetry of objects is the most fundamental a priori constraint used by the human visual system in recovering 3D shapes of the objects (Li et al., 2009, 2011). 3D mirror symmetry is easy to verify in 3D representations: pairs of mirror symmetrical points form parallel lines segments that are bisected by the symmetry plane. However, a 2D perspective image of a 3D mirror-symmetrical shape is almost never symmetrical. How is the visual system able to establish 3D symmetry correspondence in 2D asymmetrical images? Similarly to the other two correspondence problems, solving symmetry correspondence in an image is an ill-posed problem because (1) a given edge can have many possible correspondences; (2) any two edges can have infinitely many spurious 3D symmetrical interpretations (Sawada et al., 2011). In this study, we show which a priori constraints have to be used in order to correctly solve the symmetry correspondence problem. The solution begins with solving figure-ground organization (FGO) problem in 3D and 2D representations. This is done based on coarse information provided by a pair of images obtained by a stereoscopic camera. The 3D FGO is used to estimate the plane of symmetry of the 3D object, assuming that this plane is orthogonal to the ground plane, and the 2D vanishing point representing the 3D symmetry in perspective images. We then extract edges within regions of the 2D image representing individual objects (figures). Finally, we detect pairs of symmetric curves by evaluating their (i) relation to the vanishing point, (ii) relative 2D orientation and (ii) relative distance. We will illustrate this model with real images of real objects.
Meeting abstract presented at VSS 2013
This PDF is available to Subscribers Only