Machines' perception of space

Abstract

Architectural design is highly dependent on the architect's understanding of space. However, in the era of digital revolution, when efficiency and economy are the major concerns in most industrial fields, whether a computer can gain human-like understanding to read and operate space and assist with its design and analysis remains a question. This thesis focuses on the geometrical aspects of spatial awareness. Machine systems that have similar behaviors to humans' perceptions of space in geometric aspects will be developed employing techniques such as isovist and machine learning, and trained with open-sourced datasets, self-generated datasets or crowdsourced datasets. The proposed systems simulate behaviors including space composition classification, space scene classification, 3D reconstruction of space, space rating and algebraic operations of space. These aspects cover topics ranging from pure geometrical understandings to semantic reasoning and emotional feelings of space. The proposed systems are examined in two ways. Firstly, they are applied to a real-time space evaluation modeling interface, which gives a user instant insights about the scene being constructed; Secondly, they are also undertaken in the spatial analysis of existing architectural designs, namely small designs by Mies van der Rohe and Aldo van Eyck. The case studies conducted validate that this methodology works well in understanding local spatial conditions, and that it can be helpful either as a design aid tool or in spatial analysis.