Bridging the Synthetic-Real Gap: Supervised Domain Adaptation for Robust Spacecraft 6-DoF Pose Estimation

The monocular vision-based 6-DoF pose estimation involves deducing the rotation and translation data of a target spacecraft from 2D images captured from the chaser spacecraft. Accurate estimation of the 6-DoF pose from 2D images is particularly challenging in the space environment due to illumination changes, specular reflections, limited texture, and significant variations in the apparent size of the target caused by changes in range during approach [5]. Early spacecraft pose estimation methods relied primarily on geometric computer vision techniques such as edge matching, template alignment, and photogrammetry-based measurements [6]. Although effective in controlled conditions, these methods are sensitive to noise and environmental variations, which limit their robustness in real on-orbit imagery. With the rise of deep learning (DL), fully end-to-end networks have been explored, mapping raw images directly to pose parameters [7, 8]. Such approaches can implicitly learn complex visual cues, but often require vast amounts of labelled data and can be less interpretable or adaptable to new spacecraft geometries. Alternate approaches to end-to-end approaches are hybrid modular approaches to estimate spacecraft poses (Figure 1). These methods combine data-driven feature extraction with geometric model-based solvers, exploiting the strengths of both paradigms [9, 10]. In a typical pipeline, the process begins with spacecraft localization, where a deep learning (DL) object detection model predicts a bounding box enclosing the target spacecraft (see Figure 1(a)).