Improving Local Symmetry Estimations in RGB-D Images by Fitting Superquadrics

Abstract

[Abstract] Real-time manipulation tasks rely on finding good candidates for apprehension points which, in turn, usually requires the computation of local symmetries. When RGB-D images are used as input information, these local symmetries can be deduced from segmenting these images and computing geometric moments for each cluster of points. This approach gives a rough approximation because it does not take into account that the considered points lie on a surface. In this paper, to improve the quality of the symmetry estimations, we propose a simple refinement process that takes as input the estimation obtained using moments and then fits a superquadric to the considered set of points. We evaluate our approach on data collected using a Microsoft's Kinect 2 sensor. The obtained experimental results demonstrate the efficacy of the proposed approach.