Photorealistic Image Generation for Satellite Pose Estimation Using Generative Adversarial Networks

In autonomous satellite servicing operations, pose estimation is an integral process to guide the servicing satellite for rendezvous and capture of the satellite to be serviced. Convolutional Neural Network CNN-based methods show promise in satellite pose estimation. In order to train CNNs for pose estimation, sufficient quantity and quality of real training imagery that is labelled with detailed pose data are required. Such images are either unavailable or very costly to produce, often forcing augmentation using computer-generated or synthetic image datasets. In order to enable CNN-based pose estimators to fulfill their robust and efficient potential, one may draw from the distribution-matching ability of the Generative Adversarial Network GAN to modify an existing training dataset of synthetic imagery based on the characteristics of markedly fewer real images. This research focuses on the Cycle-Consistent GAN CycleGAN architecture for its strength in such style transfer tasks. Both a geometrically simple proof-of-concept object and the on-orbit images of a small satellite are employed for photorealistic image generation using CycleGAN and training of a simple CNN pose estimator. Resulting improvement to real image pose estimation accuracy of this CNN when trained on such photorealistic imagery vice synthetic imagery provides valuable insight to future applications of the implementation of CycleGAN for such training data generation.