Carnegie Mellon University
Browse

The 3D Form Analysis of Regional Architecture Using Deep Learning: A Case Study of Wood Churches from the Carpathian Mountains

Download (7.61 MB)
thesis
posted on 2022-10-25, 20:05 authored by Michael HaseyMichael Hasey

Recent research into 3-D architectural and urban form analysis using deep learning (DL) methods has shown potential to extract and identify intrinsic features from large collections of building data and draw novel insights into how our built environment is configured and interrelated. In order to explore these new capabilities, this thesis offers a detailed case study of a critical engagement with building data and DL techniques for the purposes of architectural-historical form analysis. For a case study, it documents the creation of a custom dataset of 3-D meshes of 313 wooden churches located within and near the Carpathian Mountain regions of Ukraine using photographic data and 3-D reconstruction techniques. Though the subject of ongoing scholarly interest, the complex regional style variations of these churches has made it difficult to establish clear stylistic rules that govern both their differences and similarities. As a result, this lack of consensus has led to conflicting findings and ongoing contention amongst researchers. Given this challenge, this thesis attempts to demonstrate how recent statistic based rather than traditional heuristic approaches might provide an alternate way to obtain new insight into this issue. For example, by offering a way to search for and identify both complex and nuanced form relationships and patterns among hundreds of churches at once within a single diagram. This includes detecting where stylistic overlap occurs, identifying groups of hybrid-styled churches, uncovering endemic micro styles, and revealing broad form patterns that illustrate how architectural features gradually morph across hundreds of buildings. The

thesis thus offers a path for DL-based form analysis techniques to be put in conversation with traditional architectural studies, helping identify strengths and weaknesses, as well as opportunities for hybrid and cross-methodological architectural-historical analyses. Indirectly, this thesis also demonstrates an alternate method of

architectural documentation through a combination of 3D building reconstruction from sparse imagery and architectural style encoding using DL-methods. Considering the threat to Ukrainian culture given the current war with Russia, digitally preserving both the individual churches themselves through 3D reconstruction and the rules that define their styles through DL-methods, has the potential to document and protect this important and distinctive part of Ukrainian architectural folk heritage. 

History

Date

2022-08-15

Degree Type

  • Master's Thesis

Department

  • Architecture

Degree Name

  • Master of Science in Computational Design (MSCD)

Advisor(s)

Daniel Cardoso Llach Jinmo Rhee