CS Events Monthly View
PhD DefenseGraph-Representation Learning for Human-Centered Analysis of Building Layouts |
|
||
Tuesday, December 22, 2020, 12:00pm - 02:00pm |
|||
Speaker: Vahid Azizi
Location : Remote via Webex
Committee:
Prof. Mubbasir Kapadia (chair)
Prof. Mridul Aanjaneya
Prof. Gerard de Melo
Prof. M. Brandon Haworth
Event Type: PhD Defense
Abstract: Floorplans are a standard representation of building layouts. Computer-Aided Design (CAD) applications and existing Building Information Modeling tools rely on simple floorplan representations which are not amenable to automation (e.g., generative design), and do not account for how people inhabit and occupy the space -- two key challenges which must be addressed for intelligent human-aware building design. This thesis addresses these challenges by exploring the use of graph representation learning techniques to implicitly encode the latent state of floorplan configurations, which is more amenable to automation. Specifically, we use graphs as intermediate representation of floorplans. Rooms are nodes and edges indicate a connection between adjacent rooms, either through a door or passageway. The graphs are annotated with a variety of attributes which characterize the semantic, geometric, and dynamic properties of the floorplan with respect to human-centered criteria. To address the variation in dimensionality of graphs, we utilize an intermediate sequential representation (generated by random walks) which allow us to encode the graphical structure in a fixed-dimensional representation. We propose the use of RNN-based variational autoencoder architectures to embed attribute floorplans. We enhance graph-based representations of floorplans with human occupancy and behavior attributes which are extracted by statically analysing the floorplan geometry, and by running human behavior simulations on large datasets of real and procedurally generated synthetic floorplans. We explore the potential of these representations on a variety of tasks including finding semantically similar floorplans, floorplan optimization, and generative design. Our approach and techniques are extensively evaluated through a series of quantitative experiments, and user studies with expert architects to validate our findings. The qualitative, quantitative and user-study evaluations show that our embedding framework produces meaningful and accurate vector representations for floorplans. Our models and associated datasets have been made publicly available to encourage adoption and spark future research in the burgeoning research area of intelligent human-aware building design.
: