FRI-Net: Floorplan Reconstruction via Room-wise Implicit Representation

TL;DR

FRI-Net addresses floorplan reconstruction from 3D point clouds by modeling each room as an implicit function guided by line-based primitives. It uses a DETR-based room encoder and a room-wise decoder that predicts occupancy through a three-stage line pipeline (line prediction, grouping, shape assembly), trained with a bipartite matching loss and a staged regularization strategy that first emphasizes axis-aligned lines and then diagonal features. The approach yields state-of-the-art results on Structured3D and strong performance on SceneCAD, with ablations confirming the value of height-informed pre-processing and stagewise training for learning global room geometry. By integrating geometric priors into a differentiable, global-shape-aware representation, FRINet offers robust floorplan reconstruction that better handles noise and missing data than prior corner- or box-based methods.

Abstract

In this paper, we introduce a novel method called FRI-Net for 2D floorplan reconstruction from 3D point cloud. Existing methods typically rely on corner regression or box regression, which lack consideration for the global shapes of rooms. To address these issues, we propose a novel approach using a room-wise implicit representation with structural regularization to characterize the shapes of rooms in floorplans. By incorporating geometric priors of room layouts in floorplans into our training strategy, the generated room polygons are more geometrically regular. We have conducted experiments on two challenging datasets, Structured3D and SceneCAD. Our method demonstrates improved performance compared to state-of-the-art methods, validating the effectiveness of our proposed representation for floorplan reconstruction.

Figures (6)

• Figure 1: Floorplan representation. A floorplan can be represented as an assembly of room polygons, with each polygon modeled through a room-wise implicit representation. The room-wise implicit representation can be constructed using a set of hyper-lines.
• Figure 2: Overview. (a) Given the input 3D point cloud, we use the pre-processing module to obtain the input image by considering the density of the floorplan and the height of the extracted wall. (b) For the input image, we use the room-wise encoder to output several room feature codes, each corresponding to a latent room representation; (c) For each room code, we use the room-wise decoder to output the corresponding room polygon. Multiple polygons are merged to form the final vectorized floorplan.
• Figure 3: Room-wise decoder. The room-wise decoder takes the room feature code $F$ and $n$ query points $X$ as inputs and outputs the occupancy $S$ of these points, indicating whether the points are inside or outside the corresponding room.
• Figure 4: Qualitative evaluations on Structured3D zheng2020structured3d.
• Figure 5: Qualitative evaluations on SceneCAD avetisyan2020scenecad.