ABSTRACT

Recent research into architectural form analysis using deep learning (DL) methods has shown potential to identify features from large collections of building data, shedding new light into formal aspects of our built environment. As these methods begin to enter architectural, urban, and policy design contexts, it becomes important to develop critical approaches to employing them. In this paper, we document and reflect upon our efforts to create a custom dataset of 3-D models of 331 wooden churches located within the Carpathian Mountains of Eastern Europe, and to use DL methods to explore this dataset with the goal of revealing unexpected formal traits and advancing architectural scholarship on this subject. While existing scholarship groups them into four distinct stylistic categories, our analysis reveals stylistic overlaps, previously undetected micro styles, and shared architectural features. We posit the resulting analyses as an example of an ‘architectural distant reading’ that enriches our understanding of this architectural typology through an unprecedentedly detailed portrait of its formal characteristics based on a large architectural dataset. Crucially, drawing on recent developments in critical data and algorithm studies, we show how the dataset construction and subsequent analyses, and their results, were shaped by slow, manual data curation processes, methodological constraints, subjective decisions, and engagements with archives, domain experts. We thus illustrate how DL techniques might be contextualized for architectural studies in relation to other modes of knowledge and labour, and offer a detailed case study of state-of-the-art computational methods enriching established approaches to architectural form and historical analysis.

摘要

最近使用深度学习 (DL) 方法进行建筑形式分析的研究显示出从大量建筑数据中识别特征的潜力，为我们建筑环境的形式方面提供了新的视角。随着这些方法开始进入建筑、城市和政策设计环境，开发使用它们的关键方法变得很重要。在本文中，我们记录并反思了我们为东欧喀尔巴阡山脉 331 座木制教堂创建 3D 模型的自定义数据集所做的努力，并使用 DL 方法探索该数据集，目的是揭示意想不到的形式特征并推进该主题的建筑学术。虽然现有的奖学金将它们分为四个不同的风格类别，但我们的分析揭示了风格重叠，以前未被发现的微观风格和共享的架构特征。我们将所得分析作为“建筑远读”的一个例子，通过基于大型建筑数据集对其形式特征进行前所未有的详细描述，丰富了我们对这种建筑类型学的理解。至关重要的是，利用关键数据和算法研究的最新进展，我们展示了数据集构建和后续分析及其结果是如何通过缓慢的手动数据管理过程、方法论限制、主观决策以及档案馆、领域专家的参与而形成的。因此，我们说明了如何将深度学习技术与其他知识和劳动模式相关的建筑研究结合起来，