Indoor functional subspace division from point clouds based on graph neural network

Abstract

Indoor scenes are closely related to human life and contain rich geometric and semantic information. Dividing indoor spaces from data is crucial for multiple applications, such as navigation and digital twins. However, achieving this task is challenging. Traditional indoor space division methods only represent buildings, floors, and rooms to some extent, lacking semantic descriptions of indoor elements and their spatial arrangements. To divide an indoor space more finely, a novel indoor space subdivision method is proposed in this paper. Our method leverages the flexible space subdivision framework (FSS) to categorize indoor space into free navigation subspace, object subspace, and functional subspace. To define functional subspaces, we present a taxonomy for the spatial layout patterns of common indoor elements (like tables, chairs, ceilings, walls, floors, etc.). Then, we introduce scene graphs to represent indoor 3D scenes, where each node represents an indoor element and each edge encodes the spatial relationship between the elements. Finally, a node classification network is proposed to segment indoor scene into subspaces and predicts their (semantic) functions. We select 9 buildings of Matterport3D and 6 areas in S3DIS and merge them to form our dataset for training and testing our method. Experiments yield good results with up to 90.42% accuracy and 85.28% F1-scores in overall space subdivision. Moreover, compared with the various graph node classification networks, our method has achieved the best performance in indoor space subdivisions.