CNN-BASED INITIAL LOCALIZATION IMPROVED BY DATA AUGMENTATION

摘要

Abstract. Image-based localization or camera re-localization is a fundamental task in computer vision and mandatory in the fields of navigation for robotics and autonomous driving or for virtual and augmented reality. Such image pose regression in 6 Degrees of Freedom (DoF) is recently solved by Convolutional Neural Networks (CNNs). However, already well-established methods based on feature matching still score higher accuracies so far. Therefore, we want to investigate how data augmentation could further improve CNN-based pose regression. Data augmentation is a valuable technique to boost performance on training based methods and wide spread in the computer vision community. Our aim in this paper is to show the benefit of data augmentation for pose regression by CNNs. For this purpose images are rendered from a 3D model of the actual test environment. This model again is generated by the original training data set, whereas no additional information nor data is required. Furthermore we introduce different training sets composed of rendered and real images. It is shown that the enhanced training of CNNs by utilizing 3D models of the environment improves the image localization accuracy. The accuracy of pose regression could be improved up to 69.37 % for the position component and 61.61 % for the rotation component on our investigated data set.