A multilayer feedforward network for model estimation from range data

Abstract

A novel neural architecture aimed to estimate superquadrics parameters form range data is presented. The network topology is designed to model and compute the inside-outside function of an undeformed superquadric in whatever attitude, starting from the (x,y,z) data triples. The network has been trained using backpropagation, and the weights arrangement, after training, represents a robust estimate of the superquadric parameters. The architectural approach is general, it can be extended to other geometric primitives for part-based object recognition, and performs faster than classical model fitting techniques. Detailed explanation of the theoretical approach, along with some experiments with real data, are reported.