A dataset of domestic environment room adjacency scene graphs

In this paper, we present a ground truth dataset of single-layer scene graphs representing the adjacency of rooms in 50 real-world domestic scenes from the Habitat-Matterport 3D Semantics dataset (HM3DSem). Our goal is to facilitate further research into how the underlying structure of domestic environments can be modelled. Domestic environments exhibit salient patterns that, if modelled, have the potential to improve a robot’s spatial understanding of such environments. These patterns are inherently graphical and hierarchical in nature, making Scene Graphs an ideal representation for modelling them. Scene Graphs capture the entities in a scene and the relationships between them. The dataset proposed in this paper augments the data already available in HM3DSem by providing labels for and interconnections between rooms, thereby providing a more holistic representation of the scenes. As such, the resulting dataset is a comprehensive resource for researchers in the domain. We make both the data and the software pipeline used to construct it available for use.