用于单幅航空图像高度估计的门控特征集合

从单张图片进行高度估计，严格地说，是一个不适定的问题。然而，最近的研究表明，从图像统计数据中学习到高度信息的映射既可能又可行。尽管这一领域近期付出了很多努力，但如何学习能够保留细节形状的特征，例如物体的边界和轮廓，仍然是一个未解决的问题。在这项工作中，我们提出了一个渐进式学习网络，以从粗糙到细致的方式从单张航拍图像中估计高度信息。特别地，我们引入了一个门控特征聚合模块，有效地结合了低层次和高层次的特征。所提出的方法在三个公共数据集上进行了验证，包括Vaihingen数据集、Potsdam数据集和DFC2019数据集。定量和定性的实验结果均显示，与四种相关的高度估计方法相比，所提出的方法能够从单张航拍图像中获得更准确的高度估计，特别是在保留物体边界和轮廓的能力上表现更佳。Height estimation from a single image is, strictly speaking, an ill-posed problem. However, recent studies have demonstrated that learning a mapping of height information from image statistics is both possible and feasible. Despite the considerable recent efforts devoted to this field, how to learn features that preserve detailed shapes—such as object boundaries and contours—remains an unsolved problem. In this work, we propose a progressive learning network for height estimation from single aerial images in a coarse-to-fine manner. Specifically, we introduce a gated feature aggregation module that effectively fuses low-level and high-level features. The proposed method is validated on three public datasets, namely the Vaihingen Dataset, Potsdam Dataset, and DFC2019 Dataset. Both quantitative and qualitative experimental results demonstrate that, compared with four relevant height estimation methods, the proposed method achieves more accurate height estimation from single aerial images, especially in its superior ability to preserve object boundaries and contours.