CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction

TL;DR

To address fragility of corner-based floorplan reconstruction under noise and occlusion, this paper introduces CAGE, a Continuity-Aware edGE network that reconstructs vector floorplans directly from density maps. It adopts an edge-centric representation, predicting directed, geometrically continuous edges via a dual-query transformer decoder augmented with denoising to stabilize training and refine geometry. Edge-based modeling yields watertight, topologically valid layouts and demonstrates state-of-the-art performance on Structured3D (Room F1 99.1%, Corner F1 91.7%, Angle F1 89.3%) and SceneCAD, with strong cross-dataset generalization. The method requires no heavy post-processing and maintains efficient inference, validating its practicality for CAD/BIM workflows and robotics.

Abstract

We present CAGE (Continuity-Aware edGE) network, a robust framework for reconstructing vector floorplans directly from point-cloud density maps. Traditional corner-based polygon representations are highly sensitive to noise and incomplete observations, often resulting in fragmented or implausible layouts.Recent line grouping methods leverage structural cues to improve robustness but still struggle to recover fine geometric details. To address these limitations,we propose a native edge-centric formulation, modeling each wall segment as a directed, geometrically continuous edge. This representation enables inference of coherent floorplan structures, ensuring watertight, topologically valid room boundaries while improving robustness and reducing artifacts. Towards this design, we develop a dual-query transformer decoder that integrates perturbed and latent queries within a denoising framework, which not only stabilizes optimization but also accelerates convergence. Extensive experiments on Structured3D and SceneCAD show that CAGE achieves state-of-the-art performance, with F1 scores of 99.1% (rooms), 91.7% (corners), and 89.3% (angles). The method also demonstrates strong cross-dataset generalization, underscoring the efficacy of our architectural innovations. Code and pretrained models are available on our project page: https://github.com/ee-Liu/CAGE.git.

Figures (16)

• Figure 1: Floorplan reconstruction using our CAGE network. Given an input point cloud (a), we project it into a 2D density map (b), predict two-level edge sequences (c), and reconstruct a vector floorplan by intersecting the predicted edges (d). Note that, our edge-based formulation enables the recovery of regular, topologically valid polygons even in severely occluded regions (highlighted with red circles). See Figure \ref{['fig:qual_structured3d']} for comparisons with corner-based and other SOTA methods.
• Figure 2: Edge-based Floorplan Representation. (a) Corner-based: polygons defined by sequential vertices; (b) Edge-based: walls represented as directed edges with geometric continuity; (c) Tokenization: polygons represented as sequences of edge tokens at two levels; (d) Density map generated from point cloud; (e) Ground-truth floorplan. While edge and corner representations are mathematically dual, the edge-based formulation offers greater robustness to noise and occlusion (highlighted in red).
• Figure 2: Quantitative comparison on the SceneCAD validation set scenecad. Results for prior methods are reported from roomformerfrinetpolydiffuse.
• Figure 3: Illustration of edge-to-polygon conversion. (a) Input edge sequence; (b) Four types of edge intersections based on geometric proximity; (c) Final reconstructed polygon.
• Figure 4: Architecture of the CAGE network. (a) The input point cloud is projected into a density map and processed by an image backbone and transformer encoder. (b) The transformer decoder receives two types of queries (perturbed and latent) and predicts edges defined by two endpoints, which may not correspond to polygon vertices. (c) A feed-forward network assigns class labels, with polygon matching for supervision. (d) Detailed architecture of decoder, showing the progressive refinement of edge queries with the incorporation of the novel designs of perturbation and denoising.