Floorplan embedding with latent semantics and human behavior annotations

Floorplans provide top-view representations of buildings that highlight key relationships between spaces and building components. In the last few decades, different approaches have been proposed to compare and catalogue different floorplans for design exploration purposes. Some approaches have considered floorplans as images, while others represented them as graphs. However, both image and graph-based approaches have failed to extract and utilize essential low-level space semantics and structural features. Further, they do not encode information about space utilization determined by people movement and activities in space, which are critical to analyze a building layout. To address these issues, we use deep learning techniques to develop a floorplan embedding - a latent representations of floorplans, which encodes multiple features. Specifically, we propose a novel framework that uses an attributed graph as an intermediate representation to encode space semantics, structural information and crowd behavioral features. We train Long Short-Term Memory (LSTM) autoencoders to represent these graphs as vectors in a continuous space. In addition, we contribute a floorplan dataset augmented with semantic and simulation-generated behavioral features. These representations spark new opportunities for next-gen design applications like clustering, design exploration tools and recommendations. Three different use cases are studied to show the performance of this method.