Projeto de uma rede neural SpaceYNet voltada a obter estimação de pose e profundidade em robôs móveis

Abstract

One of the main needs when working with neural networks is a good quality of training data. With that in mind, we carried out a project to create a database for training neural networks to predict 6-DoF pose. The database contains RGB and depth images, related to properly labeled odometry data. The database was then used to train SpaceYNet, inspired by the GoogLeNet and PoseNet network, which aims to regress robot position data in the environment. Comparing the network with its inspirations shows not only an improvement in functionality, but also a promising idea to be invested in.