Segmentation and Recovery of Superquadric Models using Convolutional Neural Networks

In this paper we address the problem of representing 3D visual data with\nparameterized volumetric shape primitives. Specifically, we present a\n(two-stage) approach built around convolutional neural networks (CNNs) capable\nof segmenting complex depth scenes into the simpler geometric structures that\ncan be represented with superquadric models. In the first stage, our approach\nuses a Mask RCNN model to identify superquadric-like structures in depth scenes\nand then fits superquadric models to the segmented structures using a specially\ndesigned CNN regressor. Using our approach we are able to describe complex\nstructures with a small number of interpretable parameters. We evaluated the\nproposed approach on synthetic as well as real-world depth data and show that\nour solution does not only result in competitive performance in comparison to\nthe state-of-the-art, but is able to decompose scenes into a number of\nsuperquadric models at a fraction of the time required by competing approaches.\nWe make all data and models used in the paper available from\nhttps://lmi.fe.uni-lj.si/en/research/resources/sq-seg.\n