Render-to-real image dataset and CNN pose estimation for down-link restricted spacecraft missions

In an environment of escalating usage of Low Earth Orbit, the active remediation of debris is an increasingly necessary capability. Computer vision pose estimation is a core competency of active debris remediation but state-of-the-art pose estimation methods continue to grow in size and complexity. For bandwidth limited and edge computing cases, smaller networks are more feasibly implemented. A 16,845 synthetic image dataset, applicable to the upcoming JAXA CRD2 project, is rendered and a small pose estimation network is constructed and trained on the dataset. The network is then quantized, reducing the memory requirement by a factor of 8x to a theoretical size of 5.5 MB. The 5.5 MB network demonstrates sufficient accuracy in both single image pose and motion prediction tasks when compared to the full precision 32 bit network.