3D Remote Scene Reconstruction via Graph Convolution
Xing Li, Mingyu Sun, Qiaofeng Ou, Yan Mo, Sikun Liu, Zhibo Rao · 2024
3D urban scene reconstruction is a significant remote sensing topic that provides enriched context on 3D spatial information. However, obtaining a watertight, lightweight, and detailed mesh object from remote sensing images is challenging due to sparse point clouds and inaccurate normals. This paper introduces a 3D mesh reconstruction framework to recover 3D scene information from pairwise remote sensing images, comprising three main stages. Firstly, our previous multi-task network (BGA-Net) is employed to predict the disparity maps from pairwise remote sensing images. Next, we utilize camera parameters and interpolation to reconstruct and fill point clouds, mitigating sparse point clouds (e.g., building edges or unsmooth areas). Finally, a graph convolution network (GCN) is adopted to recover the mesh object from the filled point clouds, reducing the dependency on precise normals. Extensive experiments demonstrate that our framework can qualitatively produce watertight, lightweight, and detailed mesh models.