Learning Implicit Neural Representation for Satellite Object Mesh Reconstruction

Constructing a surface representation from the sparse point cloud of satellite is an important task for on-orbit services such as docking and maintenance. In related studies on reconstruction clouds, implicit neural representations have gained popularity in learning-based 3D object reconstruction. When aiming with more complicated geometry larger intra-class variance, existing approaches cannot perform well. To solve above contradictions make effective use representations, we built NASA3D dataset containing watertight meshes, occupancy values, corresponding points by using models NASA’s official website. On basis NASA3D, propose novel network called GONet detailed grids. By designing explicit-related Grid Occupancy Field (GOF) introducing it into GONet, compensate lack explicit supervision approaches. The GOF, together field (OF), serves supervised information learning. Learning GOF strengthens GONet’s attention to critical extraction algorithm Marching Cubes; thus, helps improve reconstructed surface’s accuracy. addition, uses same encoder decoder ConvONet but designs Adaptive Feature Aggregation (AFA) module achieve adaptive fusion planar volume features. insertion AFA allows obtained features incorporate geometric volumetric information. Both visualization quantitative experimental results demonstrate that our could handle work outperform state-of-the-art methods significant margin. With mesh, achieves 5.507 CD-L1, 0.8821 F-score, 68.86% IoU, which equal gains 1.377, 0.0466, 3.59% over previous respectively.