Illumination-Aware 3D Gaussian Splatting for Spacecraft Reconstruction and Scene Composition

Abstract

Radiance fields have recently emerged as a powerful tool for learning continuous, three-dimensional representations of complex scenes from collections of discrete image observations. Neural Radiance Fields and 3D Gaussian Splatting have demonstrated high-quality novel view synthesis across a wide range of terrestrial applications; however, their use in spacecraft modeling introduces unique challenges that cannot be fully represented with standard static-scene formulations. Spacecraft are often observed under variable illumination conditions, with large changes in sun angle and stark shadows. These conditions make it difficult for conventional radiance field models to separate persistent spacecraft structure from lighting-dependent appearance, limiting traditional implementations’ usefulness for creating accurate reconstructions for spacecraft inspection, mapping, and proximity operations.

This thesis investigates radiance field modeling for spacecraft scenes, with a particular focuson 3D Gaussian Splatting for dynamic relighting, shadowing, and multi-object scene composition. Building on the modular radiance field training and evaluation framework Nerfstudio, this work introduces an interactive pipeline for rendering multiple independently trained radiance field models within a shared scene, develops sun-angle-conditioned relighting and shadowing extensions for spacecraft modeling, and explores training strategies intended to improve reconstruction quality and learning stability in spacecraft-imaging domains. The proposed methods are evaluated across simulated, synthetic, and real-world datasets, demonstrating their promise and practical limitations for spacecraft scene reconstruction over a spectrum of use cases. Together, these contributions move radiance fields toward more flexible, illumination-aware representations for spacecraft modeling, enabling realistic visualization and analysis under the dynamic lighting, occlusion, and relative positioning conditions of space operations.