边界框足矣：基于检测建筑上下文编码的街景图像分类
摘要
面向城市土地利用分析的街景图像分类十分困难，因为其类标签（如商业区）相比通用视觉任务（如行人与汽车）的标签具有更高的抽象层级。因此，仅使用视觉特征的分类模型往往难以取得满意性能。本文提出一种基于“检测器-编码器-分类器”框架的新方法。与基于卷积神经网络（CNN）的常规图像级模型直接采用整图视觉特征不同，该框架首先由检测器获取街景图像中建筑的边界框；随后由所提算法“CODING”（检测建筑上下文编码，Context encOding of Detected buildINGs）将其类别共现模式与布局等上下文信息编码为元数据；最后由循环神经网络（RNN）对这些边界框元数据进行分类。此外，基于现有 BIC GSV [1]，我们制作了含 19,070 张街景图像与 38,857 栋建筑、具双重标签的数据集“BEAUTY”（建筑检测与城市功能区刻画，Building dEtection And Urban funcTional-zone portraYing）。该数据集不仅可用于街景图像分类，亦可用于多类建筑检测。在“BEAUTY”上的实验表明，相较图像级 CNN 模型，所提方法在宏精度上提升 12.65%，在宏召回上提升 12%。我们的代码与数据集见 https://github.com/kyle-one/Context-Encoding-of-Detected-Buildings/
引用
@article{arxiv.2010.01305,
title = {Bounding Boxes Are All We Need: Street View Image Classification via Context Encoding of Detected Buildings},
author = {Kun Zhao and Yongkun Liu and Siyuan Hao and Shaoxing Lu and Hongbin Liu and Lijian Zhou},
journal= {arXiv preprint arXiv:2010.01305},
year = {2021}
}
备注
Figure 1 has been added, and the order of the rest of the figures continues. Figure 6 (Figure 5 of the previous version) and Figure 7 (Figure 6 of the previous version) have been modified. Figure 7, Figure 15, and Figure 16 of the previous versionhave have been removed. The structure of Section 4 has been adjusted