Deep Learning Isovist

Understanding the qualitative aspect of space is essential in architectural design. However, the development of computational design tools has lacked features to comprehend architectural quality that involves perceptual and phenomenological aspects of space. The advancement in machine learning opens up a new opportunity to understand spatial qualities as a data-driven approach and utilize the gained information to infer or derive the qualitative aspect of architectural space. This paper presents an experimental unsupervised encoding framework to learn the qualitative features of architectural space by using isovist and deep learning techniques. It combines stochastic isovist sampling with Variational Autoencoder (VAE) model and clustering method to learn and extract spatial patterns from thousands of floorplans data. The developed framework will enable the encoding of architectural spatial qualities into quantifiable features to improve the computability of spatial qualities in architectural design.