Application of Convolutional Neural Networks for Three-Dimensional Reconstruction of the Geometry of Objects in the Image

Abstract

Automatic creation of three-dimensional prototypes and digital copies of three-dimensional objects of the real world is a revolutionary innovation that is actively used today in many spheres of human activity, for example, for identification of a person for various purposes, in intelligent transport systems, in visualization systems, as well as in digital systems for design, management and monitoring. This trend has intensified at the present time, when additive technologies have become available to a wide range of users, and large-scale repositories of three-dimensional objects are becoming increasingly popular and widespread. The problem of recognizing images of objects of the surrounding world and understanding their scale and volume by two-dimensional projections is one of the most urgent and studied problems solved by computer vision methods. However, this class of tasks is quite difficult to formalize, which makes their solution time-consuming to develop and implement. The article describes the development of a software complex that reconstructs three-dimensional scenes according to their projections using neural network machine learning methods: the basics of three-dimensional reconstruction are considered, a model of the overall architecture of the hardware and software complex is proposed the architecture of the developed neural network is given, as well as the results of training and technical experiments.