AttentionBuildNet for Building Extraction from Aerial Imagery

Abstract

Extracting building footprints accurately from high-resolution aerial imagery significantly impacts an extensive range of applications such as change detection, automatic mapping, urban planning, and monitoring unauthorized land use. Automatic building extraction remains challenging due to complex structures, different textures and appearance, and small and densely connected buildings. This paper proposes a novel approach, AttentionBuildNet (ABNet), that precisely extracts building footprints and boundaries. The proposed model improves the overall feature representation by selectively focusing on important features by utilizing a convolution block attention module with the channel and spatial attention. We introduce a new unit, cross attention module, to capture multi-scale features with different dilation rates. We evaluate on Massachusetts Building Dataset, and from the results, it is clear that the proposed ABNet model surpasses all other previous methods.