Self-Distillation Adaptive Pose Estimation Method for Cross-Domain Non-Cooperative Spacecraft Based on Optical Sensor Data

Domain adaptation is crucially important for non-cooperative spacecraft pose estimation from optical sensor data in real-world orbital images. Given the scarcity of real images, synthetic images generated through rendering serve as a practical substitute. However, synthetic data fails to accurately simulate complex orbital conditions, and such domain shift degrades model performance in real-world applications. To address these challenges, this paper proposes a self-distillation unsupervised domain adaptation (UDA) framework designed to learn domain-invariant representations. Initially, the pre-trained Vision Transformer (ViT) is fine-tuned on the synthetic source domain via heatmap supervision to improve model generalizability. Subsequently, masked local inference on multiple unlabeled target domains allows the model to learn discriminative local features and spatial contextual relationships. This process is further enhanced by self-distillation, which facilitates progressive knowledge transfer from the source domain. Additionally, an optimizable shared prototype space is designed to implicitly encode semantic structures, enabling cross-domain semantic consistency alignment. Finally, experiments demonstrate that the proposed method achieves state-of-the-art accuracy on the SPEED+ benchmark dataset.