DV2BR: A Dual-View 3D Building Reconstruction Method From Aerial Images

Aiming at the problem of how to achieve higher-quality three-dimensional (3D) modelling of buildings using optical remote sensing images under the scarcity of remote sensing resources, a 3D implicit modelling network based on dual-view is proposed. In order to better capture the shape and structural features of the building, we introduce an attention mechanism in the coding stage to improve the feature extraction capability. To further improve the encoder's ability to extract global features, we also designed a multi-view feature fusion module. In the decoding stage, we introduce an implicit representation based on the Symbolic Distance Field (SDF), which allows for finer details. Sufficient experiments demonstrate that compared with previous 3D modelling methods, our proposed method provides substantial improvements in both the accuracy of the generated building model and the restoration of the details.