Style Transfer Framework Improves Satellite Pose Estimation from Synthetic Data

A novel framework for style transfer, which is aware of components and preserves structure, has been introduced by researchers to help close the synthetic-to-real divide in monocular 6D pose estimation for non-cooperative satellites. This technique utilizes weakly paired real-synthetic samples derived from calibrated real acquisitions, ArUco-based camera-pose measurements, CAD rendering, and component masks. It captures part-specific style codes from unlabeled real images and integrates them into the relevant synthetic satellite areas through mask-aligned modulation. To ensure that the generated images are suitable for further supervision, adversarial training is employed. This method tackles the issue of the limited availability of annotated real satellite images that have dependable pose labels and component-level masks.