Object-Level Space Target Reconstruction via Semantic-Guided 3D Gaussian Splatting

On-orbit servicing missions demand accurate 3D modeling of space targets, which is typically achieved by reconstructing from multi-view observations. While Standard 3D Gaussian Splatting (3DGS) delivers impressive rendering performance, it is inherently designed to model the entire visual field and consequently tends to waste representational capacity on the dark, uninformative background regions typical of space images. To overcome this limitation, we present a semantic-driven 3D Gaussian reconstruction framework. Our approach shifts away from holistic scene modeling by integrating multi-view target masks into the optimization process, thereby forging a representation that is explicitly object-aware. To improve target-region consistency under weak texture, we augment the model with a boundary-aware depth regularization term that encourages locally smooth rendered depth while preserving visible boundaries. Validated against a synthetic spacecraft dataset generated in Unity, our method demonstrates cleaner target-centric renderings, more accurate mask agreement, and fewer background artifacts relative to conventional 3DGS baselines.