Semantic-aware room-level indoor modeling from point clouds

This paper introduces a framework for reconstructing fine-grained room-level models from indoor point clouds. The motivation behind our method stems from the consistent floorwise appearance of building shapes in urban buildings along the vertical direction. To this end, each floor's points are horizontally sliced to obtain a representative cross-section, from which the linear primitives are detected and enhanced. These linear primitives help to divide the entire space into non-overlapping connected faces with shared edges. These faces are then classified as indoor or outdoor categories by solving a binary energy minimization formulation. The indoor faces are further grouped into each individual rooms with the support of the room semantic map. By propagating and tracing each room's contour, 2D floor plan can be generated in a semantic-aware manner. These generated 2D floor plans are vertically stretched to match the heights of their respective rooms. Experimental results on six complex scenes from the S3DIS dataset, which encompass both linear and nonlinear shapes, demonstrate that our created room models exhibit accurate geometry, correct topology, and rich semantics. The source code of our room-level modeling algorithm is available at https://github.com/indoormodeling/indoor-modeling.