Height Estimation Based on Semantic Segmentation

Buildings are crucial for urban development, and building extraction and height estimation are two essential components in the process of architectural reconstruction. Nowadays, the use of instance segmentation methods to extract buildings from remote sensing imagery has become quite mature. However, accurately estimating the heights of complex buildings from images remains highly challenging. To address the task of building height estimation, this paper attempts to use semantic segmentation methods. Firstly, the height label quantification is simplified into a semantic segmentation task. Then, the segmentation results from different models are integrated through voting to improve the segmentation accuracy of buildings. Finally, the average heights predicted by different models are obtained based on the positions containing building regions to obtain the final height estimation. In the DFC2023 competition, our team ranked first in the height estimation task, sixth in the building extraction task, and third in the overall ranking.