基于多角度图像的建筑物三维重建

Building 3D Reconstruction Based on Multi-angle Images

如今,建筑物的三维重建不仅是一个概念,在遥感领域、城市建设发挥着不可替代的作用。因为传统重建方法对场景的低纹理、高光和反射区域使密集匹配变得困难,导致不完整的重建。针对上述问题,提出了一个自适应聚合的递归多视图立体网络。该网络具有两个自适应聚合模块。一是视图内聚合模块,用于鲁棒的特征提取,其中上下文感知特征自适应地聚合到具有不同纹理丰富度的多个尺度和区域。一是视图间聚合模块,用于多视图代价体聚合,其通过在良好匹配的视图对上分配更高的权重来克服复杂场景中不同遮挡的困难。利用DTU数据集对提出的框架进行实验,与现有方法相比,该框架能够获得更准确的深度图,生成更密集和完整的点云。

Nowadays,3D reconstruction of buildings is not only a concept,but plays an irreplaceable role in the field of remote sensing and urban construction.Because traditional reconstruction methods make dense matching difficult for the low-texture,high-light and reflective areas of the scene,leading to incomplete reconstruction.For the above problems,a recursive multi-view stereo network with adaptive aggregation is proposed.The network has two adaptive aggregation modules.One is an intra-view aggregation module for robust feature extraction,where context-aware features are adaptively aggregated to multiple scales and regions with different texture richness.One is an inter-view aggregation module for multi-view cost aggregation,which overcomes the difficulty of occlusion variation in complex scenes by assigning higher weights on well-matched view pairs.The proposed framework is experimented using DTU dataset,and it is able to obtain more accurate depth maps and generate more dense and complete point clouds compared with existing methods.