Recovery of superquadric parameters from depth images using deep learningOblak, Tim (Avtor) Solina, Franc (Mentor) Roth, Peter M. (Komentor) superquadricsparametric modelsreconstruction3Ddeep learningconvolutional neural networksCNNparameter recoveryReconstruction of 3D space from 2D image data has always been a significant challenge in the field of computer vision. Simple geometric entities are used to describe larger, more complex objects or entire scenes. This representation of the environment allows an autonomous agent to manipulate and interact with it's surroundings. Superquadrics are parametric models, able to describe a wide array of 3D objects using only a few parameters, which makes them a suitable representation in such tasks. In this work, we explore the possibility of using deep learning techniques to successfully recover parameters of a single superquadric from depth images. We present a new framework, which enables us to train deep learning models able to interpret the ambiguous nature of superquadrics in general position. We propose multiple loss functions for usage in supervised and unsupervised learning scenarios. On a synthetic depth image dataset, our best CNN regression model achieves an IoU accuracy of 95% and a speedup of a factor of 240 compared to the classic iterative recovery method.20202020-09-25 10:40:02Magistrsko delo/naloga120770sl