Semantic Tree-Based 3D Scene Model Recognition

Abstract

3D scene recognition is important for many applications including robotics, autonomous driving cars, augmented reality (AR), virtual reality (VR), 3D movie and game production. A lot of semantic information (i.e. objects, object parts and object groups) is existing in 3D scene models. To significantly improve 3D scene recognition accuracy, we incorporate such semantic information into the recognition process by building a semantic scene tree and propose a deep random field (DRF) model-based semantic 3D scene recognition approach. Experiments demonstrate that the semantic approach can effectively capture semantic information of 3D scene models, accurately measure their similarities, and therefore greatly enhance the recognition performance. Code, data and experimental results can be found on the project homepage.