Abstract
The proliferation of indoor maps is limited by the manual process of generalizing floor plans. Previous attempts at automating similar processes use rasterization. With Graph Neural Networks (GNN) it is now possible to rely on the inherent structures in CAD drawings and thus avoid rasterization. Localization of doors is a core component in floor plan generalization. We show how state-of-the-art GNNs can be used to classify graph nodes, extracted from CAD primitives, as door or non-door. Generalization is represented by the creation of bounding boxes from the labelled graph nodes. Our baseline is a standard raster-based bounding box detector. To support further development of graph-based methods and comparison with raster-based methods, we publish a new dataset that consists of both image and graph-based floor plan representations. Our graph-based approach completely outperforms the Faster-RCNN baseline, which fail to locate any doors with the desired localization accuracy. Code and dataset is available at https://github.com/Chrps/MapGeneralization.
Info
Conference Paper, 2021
UN SDG Classification
DK Main Research Area
Science/Technology
