ArchitectureStyle

建筑风格分类不同于标准分类任务，因为不同风格之间有丰富的阶级间关系，例如重新诠释，复兴和地域性。在本文中，我们采用基于可变形零件的模型 (DPM) 来捕获基本建筑构件的形态特征，并提出了引入概率分析并解决潜在变量模型中的多类问题的多项式潜在逻辑回归 (MLLR)。由于缺乏公开可用的数据集，我们发布了一个新的大型建筑风格数据集，其中包含25个类。对该数据集的实验表明，MLLR与标准的全局图像特征相结合，获得了最佳的分类结果。我们还为结果提供可解释的概率解释，例如单个建筑物的样式和样式关系网络，以说明类间关系。Architectural style classification differs from standard classification tasks, as there exist rich inter-class relationships between different styles, such as reinterpretation, revival, and regionalism. In this paper, we adopt the Deformable Part Model (DPM) to capture the morphological features of basic architectural components, and propose the Multinomial Latent Logistic Regression (MLLR), which introduces probabilistic analysis and addresses multi-class problems in latent variable models. Due to the lack of publicly available datasets, we release a new large-scale architectural style dataset consisting of 25 classes. Experiments conducted on this dataset demonstrate that combining MLLR with standard global image features achieves the best classification performance. We also provide interpretable probabilistic explanations for the results, such as individual building styles and style relationship networks, to illustrate the inter-class relationships.

搜集汇总

背景与挑战

背景概述

该数据集是一个大型建筑风格分类数据集，包含25个类别，专门设计用于研究建筑风格间的阶级间关系（如重新诠释、复兴和地域性）。它由上海交通大学和悉尼科技大学于2020年发布，旨在支持基于可变形零件模型和多项式潜在逻辑回归等方法的分类实验，并提供可解释的概率分析。