Three-dimensional pose from two-dimensional images: a novel approach using synergetic networks

Abstract

Neural networks have been successfully applied in many applications of machine vision. In this work, a synergetic network is used to estimate the pose of a rigid three-dimensional object. The estimation is based on a number of two-dimensional snapshots of the object with known pose. The algorithm at the base of the synergetic computer can be realised as a neural network with a two-layer topology and units that calculate dot products. In the process of constructing this network, the dimensionality of the problem is reduced dramatically from N, the number of pixels, to M, the number of prototype images. In contrast to traditional pose estimation techniques, this approach is based on appearance, rather than a detailed knowledge of shape and reflectance properties, making it flexible and amenable to situations where a detailed description of the object is not available. The algorithm is demonstrated to have fast recall times, opening the possibility of developing a real-time pose estimation system for use with robotic manipulation.