Cross-Task Mamba Network for Building Extraction and Height Estimation from Single-View Remote Sensing Images

Abstract. Simultaneous building extraction and height estimation from single-view satellite imagery via multi-task learning presents a viable solution for large-scale urban 3D reconstruction. However, balancing the weights across different tasks and reducing conflicts between them remains a challenging problem. In this paper, we propose a Mamba-CNN based network to more effectively capture the spatial distribution of buildings. Additionally, we propose a cross-task feature fusion module to facilitate information exchange between tasks. Experiments conducted on the Vaihingen dataset demonstrate significant improvements achieved by the proposed method.