Attention Based Multi-Scale Network for Height Estimation

With the rapid advancement of 3D reconstruction, accurately extracting multidimensional building information from high-resolution remote sensing images remains a key research challenge. Existing single-image height estimation methods struggle with multi-scale feature fusion due to complex terrain and building structures, leading to difficulties in detail recovery. To address this, we propose Attention based multi-scale network (AMNet) for height estimation, an encoder-decoder model integrating self-attention and multi-scale feature fusion. AHENet employs Flexible Attention Mechanisms (FAM) in the encoder to enhance feature representation, Dense Atrous Spatial Pyramid Pooling (DenseASPP) for balancing global and local information, and a Progressive Fusion Module (PFM) in the decoder to refine multi-scale feature integration. Experiments on the OSI dataset demonstrate that AHENet outperforms traditional methods and other deep learning models in remote sensing height estimation.