Machines’ Perception of Space: Employing 3D Isovist Methods and a Convolutional Neural Network in Architectural Space Classification

Abstract

Simple and common architectural elements can be combined to create complex spaces. Different spatial compositions of elements define different spatial boundaries, and each produces a unique local spatial experience to observers inside the space. Therefore an architectural style brings about a distnct spatial experience. While multple representation methods are practced in the field of architecture, there lacks a compelling way to capture and identfy spatial experiences. Describing an observer's spatial experiences quanttatively and efficiently is a challenge. In this paper, we propose a method that employs 3D isovist methods and a convolutional neural network (CNN) to achieve recognition of local spatial compositions. The case studies conducted validate that this methodology works well in capturing and identfying local spatial conditions, illustrates the patern and frequency of their appearance in designs, and indicates peculiar spatial experiences embedded in an architectural style. The case study used small designs by Mies van der Rohe and Aldo van Eyck. The contribution of this paper is threefold. First, it introduces a sampling method based on 3D Isovist that generates a 2D image that can be used to represent a 3D space from a specific observation point. Second, it employs a CNN model to extract features from the sampled images, then classifies their corresponding space. Third, it demonstrates a few case studies where this space classification method is applied to different architectural styles.