Bridging the Synthetic-to-Real Gap in Spacecraft Recognition: A Hardware-in-the-Loop Dataset

Deep learning-based spacecraft recognition is hindered by the limited availability of real in-orbit imagery, affecting model generalization. Although synthetic datasets alleviate challenges such as annotation costs and feature diversity, significant domain gaps persist. This study presents the Hardware-in-the-Loop Spacecraft Dataset (HILSD), constructed by capturing scaled satellite models under controlled space-relevant lighting and imaging conditions. HILSD comprises over 12,000 images from five structurally distinct spacecraft models, offering significantly greater diversity than prior HIL datasets. HILSD is combined with synthetic datasets to reduce domain discrepancies. Experimental results show that integrating HILSD into training pipelines substantially improves model performance on real-world imagery, particularly when combined with a core-set selection strategy that mitigates overfitting in multi-source domain learning.