Spacecraft-NeRF: High-Fidelity Reconstruction of Spacecraft by Neural Radiance Field Based Implicit Representation

Access to the appearance and geometry of spacecraft is one of the key points for conducting on-orbit services. However, without prior information, most contemporary methods still collect raw data from active sensors and then carry out a series of complex reconstruction processes. Therefore, simple and tractable methods for high-fidelity spacecraft reconstruction and rendering remain challenging in the current on-orbit services. Recently, neural radiance field-based implicit representation has demonstrated outstanding performance in a variety of reconstruction tasks. In this work, we focus on the basic needs in maintenance and fault diagnosis in on-orbit service: 1) 2-D view synthesis, 2) 3-D model reconstruction, and propose a high-fidelity reconstruction method, Spacecraft-NeRF. It masks out the complex background content in the outdoor scenario with mask images generated by the segment anything model, solves the sampling and rendering problems with the$L_{\infty }$norm construction and a small proposal MLP, and improves the reconstruction quality via a hybrid encoding strategy. Based on simulated satellites, we collected and published a dataset, Spacecraft-3D, consisting of four types of spacecraft with different surface textures and geometric structures. In this dataset, Spacecraft-NeRF shows realistic rendering performance, and the extracted 3-D mesh models properly represent the complex mechanical and geometrical structure of the spacecraft body. In comparison experiments with a series of representative NeRF-based models, Spacecraft-NeRF also outperforms most of the comparison methods in PSNR, SSIM, and LPIPS metrics.