Reflective Symmetry Detection by Rectifying Randomized Correspondences

We present a method for detecting bilateral or reflective symmetries in images. We pose the problem as an instance of a multiple model estimation problem. We build candidate symmetry models by randomly sampling minimal sets of SIFT matches. Since these symmetry models can be in non-frontal viewpoints, we rectify them, undoing the perspective effect. From the models with valid symmetric properties, we compute consensus sets by determining which SIFT matches are compatible with each symmetry model. We finally recombine these consensus sets, using a clustering algorithm. The method is able to detect single and multiple symmetries both in frontal and non fronto-parallel viewpoints, achieving state-of-the-art results.