Height Estimation from Single Aerial Imagery with a Deep Boundary-Guided Network

Extracting 3D information from single aerial image plays an important role in computer vision and remote sensing. However, due to the structural complexity of ground objects and noise introduced during the generation stage of ground truth labels, it is challenging to automatically recover the regularized height map from only one orthogonal photography. In this paper, we propose a novel deep network for estimating accurate and regularized height map from a single aerial image. The network mainly contains two sub-networks, namely the height map derivation sub-network and the boundary guidance sub-network. They are sequentially connected together, so that the corresponding boundary map can be directly calculated after the height map is obtained. We also propose a loss function suitable for semantic boundary guidance, which is similar to SSIM loss function at the edges of the ground targets. Apart from pursuing accuracy of height regression, boundary regularity constraints derived from semantic labels are also employed to form a joint metric criterion. We perform a qualitative and quantitative evaluations on ISPRS remote sensing dataset, and the result indicate that our framework improve both accuracy and regularity of estimated depth map.