End-to-end Neural Estimation of Spacecraft Pose with Intermediate Detection of Keypoints

State-of-the-art methods for estimating the pose of spacecrafts in Earth-orbit images rely on a convolutional neural network either to directly regress the spacecraft's 6D pose parameters, or to localize pre-defined keypoints that are then used to compute pose through a Perspective-n-Point solver. We study an alternative solution that uses a convolutional network to predict keypoint locations, which are in turn used by a second network to infer the spacecraft's 6D pose. This formulation retains the performance advantages of keypoint-based methods, while affording end-to-end training and faster processing. Our paper is the first to evaluate the applicability of such a method to the space domain. On the SPEED dataset, our approach achieves a mean rotation error of $$4.69^\circ $$ and a mean translation error of $$1.59\%$$ with a throughput of 31 fps. We show that computational complexity can be reduced at the cost of a minor loss in accuracy.