Multi-path Fusion Network for High-Resolution Height Estimation from a Single Orthophoto

Height estimation from a single orthophoto is essential for reconstructing 3D scene models for navigation of unmanned aerial vehicles. It is particularly challenging to recover detailed object structures under various scales. In this paper, we propose a multi-path fusion network for generating high resolution height maps while preserving scene structures well. From multi-scale features that can be extracted by an efficient recursive refinement network, we introduce a multipath feature fusion module to combine these features and exploit information on different abstraction levels efficiently. We also design a residual up-sampling block to generate high-resolution height maps that well preserve structure details. Experimental results on two public datasets, Vaihingen and Potsdam, demonstrate that our method achieves better quantitative performance compared with previous techniques. Moreover, our results are much more visually pleasing because scene structures at different scales are well preserved.