Semantic floorplan segmentation using self-constructing graph networks

Abstract

This article presents an approach for the automatic semantic segmentation of floorplan images, predicting room boundaries (walls, doors, windows) and semantic labels of room types. A multi-task network was designed to represent and learn inherent dependencies by combining a Convolutional Neural Network to generate suitable features with a Graph Convolutional Network (GCN) to capture long-range dependencies. In particular, a Self-Constructing Graph module is applied to automatically induce an input graph for the GCN. Experiments on different datasets demonstrate the superiority and effectiveness of the multi-task network compared to state-of-the-art methods. The accurate results not only allow for subsequent vectorization of the existing floorplans but also for automatic inference of layout graphs including connectivity and adjacency relations. The latter could serve as basis to automatically sample layout graphs for architectural planning and design, predict missing links for unobserved parts for as-built building models and learn important latent topological and architectonic patterns.