저널 : JICRS(Journal of Institute of Control, Robotics and Systems), Vol.25, No.8, pp.690-704, 2019. 8.
· 논문제목 : “Modelling Dynamics of Indoor Environment with 3D Scene Graphs”
· 저자 : 신동협, 김인철
· 요약 : Recent works on visual scene understanding attempt to build a scene graph. A scene graph is a rich semantic representation of the objects and their pair-wise relationships in an environment. Most of existing works have studied how to generate a 2D scene graph that represents the objects and their relationships detected from a single image. A 2D scene graph is usually a partial description of only a single state of the environment. Therefore, most of existing approaches do not address the dynamics and uncertainty of 3D real-world environments. This paper supposes both partially observable and dynamic 3D indoor environments that can be changed by agent’s actions and tries to model the environmental changes with 3D scene graphs. The proposed system attempts to estimate the correct state of a dynamic environment by combining the state recognition based on image analysis with the state prediction based on action models. Conducting lots of experiments with a large dataset collected from a virtual 3D indoor environment, AI2-THOR, we showed high performance of the proposed system.