PolyRoom: Room-aware Transformer for Floorplan Reconstruction

TL;DR

This work tackles indoor floorplan reconstruction from unstructured point clouds by producing vectorized polygon sequences. It introduces PolyRoom, a room-aware Transformer that employs a uniform sampling representation, room-aware query initialization from instance segmentation, and room-aware self-attention to progressively refine fixed-length vertex sequences for each room. The model achieves state-of-the-art performance on Structured3D and SceneCAD, with strong generalization and qualitative improvements over prior methods, while reducing memory demands via its attention design. Overall, PolyRoom advances robust, scalable floorplan reconstruction and opens avenues for more editable, topology-preserving indoor maps useful in AR/VR, robotics, and navigation.

Abstract

Reconstructing geometry and topology structures from raw unstructured data has always been an important research topic in indoor mapping research. In this paper, we aim to reconstruct the floorplan with a vectorized representation from point clouds. Despite significant advancements achieved in recent years, current methods still encounter several challenges, such as missing corners or edges, inaccuracies in corner positions or angles, self-intersecting or overlapping polygons, and potentially implausible topology. To tackle these challenges, we present PolyRoom, a room-aware Transformer that leverages uniform sampling representation, room-aware query initialization, and room-aware self-attention for floorplan reconstruction. Specifically, we adopt a uniform sampling floorplan representation to enable dense supervision during training and effective utilization of angle information. Additionally, we propose a room-aware query initialization scheme to prevent non-polygonal sequences and introduce room-aware self-attention to enhance memory efficiency and model performance. Experimental results on two widely used datasets demonstrate that PolyRoom surpasses current state-of-the-art methods both quantitatively and qualitatively. Our code is available at: https://github.com/3dv-casia/PolyRoom/.

Figures (6)

• Figure 1: Structured floorplan reconstruction by PolyRoom. Given an indoor point cloud (a), PolyRoom predicts the initial room shapes (c) from the density map (b). Then, PolyRoom refines them gradually (d) and reconstructs the final vectorized floorplan (e).
• Figure 2: Overall architecture of PolyRoom. PolyRoom consists of four main components: (a) Encoder module, (b) Decoder module, (c) Room-aware query initialization module, and (d) Floorplan extraction module. Room queries are initialized with instance segmentation. Subsequently, they are refined in the Transformer decoder layer by layer with dense supervision (red and blue boxes mark the changes). Finally, the floorplan is extracted based on vertex selection. The detailed structure of the $i$th layer in the Transformer decoder is depicted in the right part, where $F$ denotes the output of the Transformer encoder, $C_i$, $C_{i+1}$ represent content queries from different layers, while $Q_i$, $Q_{i+1}$ denote room queries from different layers.
• Figure 3: (a) Illustration of floorplan representation including the sparse corner-based (left) and our dense uniform sampling representation (right). Valid contour vertices (outlined) and corner vertices (filled) with supervision during training are colored according to the room on the left. (b) Illustration of the self-attention variants including the room-aware self-attention and vanilla self-attention. Our room-aware self-attention is a combination of intra-room and inter-room self-attention, which works among different vertices in a single room and among different rooms. And the vanilla self-attention performs on the flattened queries.
• Figure 4: Qualitative evaluations on Structured3D zheng2020structured3d. PolyRoom performs with fewer missing rooms, more correct room sequences, and improved room details.
• Figure 5: Qualitative evaluations on SceneCAD avetisyan2020scenecad.