TLDR: The CAGE (Continuity-Aware edGE) network is an end-to-end framework that reconstructs vector floorplans directly from point-cloud density maps. Unlike traditional corner-based methods, CAGE uses an edge-centric formulation, modeling wall segments as continuous edges. This approach, combined with a dual-query transformer decoder and denoising framework, significantly improves robustness to noise and incomplete data, ensuring watertight and topologically valid room boundaries. CAGE achieves state-of-the-art performance on Structured3D and SceneCAD datasets, demonstrating superior accuracy and strong cross-dataset generalization.
Reconstructing accurate and editable floorplans from 3D scans is a long-standing challenge in computer vision and robotics. These vector floorplans, which are structured 2D representations of interior spaces, are crucial for applications ranging from building management to augmented reality and autonomous navigation. However, real-world scans often suffer from noise, occlusions, and incomplete data, making precise reconstruction difficult.
Traditional methods often rely on detecting individual corners to define room polygons. While seemingly intuitive, this approach is highly sensitive to imperfections in the data. A single missed or misplaced corner can lead to fragmented or unrealistic layouts. Even more recent techniques that group lines to infer structure still struggle with capturing fine geometric details accurately.
Introducing CAGE: A Novel Edge-Centric Approach
A new research paper introduces an innovative solution called CAGE (Continuity-Aware edGE) network. This end-to-end framework rethinks how floorplans are represented and reconstructed. Instead of focusing on corners, CAGE models each wall segment as a directed, geometrically continuous edge. This fundamental shift in representation offers significant advantages.
By focusing on edges, CAGE inherently promotes the inference of coherent floorplan structures. This means it can generate watertight and topologically valid room boundaries, even when dealing with noisy or incomplete input data. The edge-centric design makes the system more robust to common real-world scanning issues and significantly reduces reconstruction artifacts.
How CAGE Works
The CAGE network takes 3D point clouds as input, first projecting them into 2D density maps. These maps are then processed by an image backbone and a transformer encoder to extract rich features. The core innovation lies in its dual-query transformer decoder. This decoder integrates two types of queries: ‘perturbed’ queries for a denoising framework and ‘latent’ queries for the final floorplan prediction.
This dual-query design is crucial for stabilizing the training process and accelerating convergence. It essentially teaches the model to recover clean and accurate edge structures from deliberately corrupted inputs, making it highly resilient to real-world noise. As the decoding progresses through multiple layers, the edge predictions are iteratively refined, transforming from coarse outlines to precise, watertight polygons.
The CAGE framework also includes a robust method for converting these predicted edge sequences into closed polygons by intelligently resolving intersections, ensuring continuity and accuracy.
Also Read:
- Mapping Indoor Spaces: A New Approach to Floorplan Reconstruction with AI
- Advancing 2D-3D Sensor Alignment for Autonomous Systems with CrossI2P
Impressive Performance and Generalization
Extensive experiments on two major indoor datasets, Structured3D and SceneCAD, demonstrate CAGE’s state-of-the-art performance. It achieves high F1 scores across various metrics, including rooms (99.1%), corners (91.7%), and angles (89.3%). Notably, CAGE outperforms previous methods, especially in fine-grained corner localization and angular estimation, without needing extensive post-processing.
Furthermore, CAGE exhibits strong cross-dataset generalization, meaning it performs well even when trained on one dataset and tested on another. This highlights the efficacy of its architectural innovations and its ability to capture fundamental structural properties of floorplans. The network is also highly efficient, matching the fastest inference speeds among leading methods.
The CAGE network represents a significant step forward in robust floorplan reconstruction, offering a more stable, accurate, and efficient solution for generating editable vector floorplans from challenging 3D scan data. You can find more details about this groundbreaking work in the research paper: CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction.
