Deep pose uncertainty learning based on a log likelihood loss for low-quality satellite images

Space images exhibit high contrast and low signal-to-noise ratios, and contain some targets of symmetry or truncated. As a result, vision-based pose estimation of noncooperative spacecraft suffers from high uncertainties. Incorporating uncertainty information along with the predicted pose becomes crucial for providing confidence in pose measurements to the filter and achieving accurate pose tracking of noncooperative spacecraft. Deep probabilistic network has been widely applied among object orientation estimation tasks with uncertainty quantification since it can learn implicit mapping relationship between RGB images and probability distribution parameters without any prior information, such as 3D models and artificial marks of the target. Existing studies either only model the probability distribution of attitude, or only split the pose uncertainty into two independent probability distributions, i.e., the matrix Fisher distribution and the multivariate Gaussian distribution. The former cannot model SE(3) (pose) uncertainty, and the latter does not consider correlation information between SO(3) (rotation matrix) and R3 (position vector), inconsistent with the fact that mutual coupling of attitude and position leads to the uncertainty of image features. To accommodate these issues, we present a deep probabilistic network that performs the relative pose estimation and model pose uncertainties. The advantage of the proposed network is that it directly models the uncertainty while avoiding the complex closed form solution for maximum likelihood estimation (MLE) of joint distribution. Specifically, the proposed network not only learns parameters of the marginal distribution on S3 (unit quaternion) and R3 (position vector) but also quantifies correlation relationship between S3 and R3 by optimizing negative log-likelihood of the Bingham-Gaussian distribution. Moreover, a simulation method as part of the computation graph is utilized to efficiently compute the non-trivial normalization constant of the Bingham-Gaussian distribution. Extensive experiments are carried out on challenging dragon-hard dataset, to validate uncertainty quantification capability of the proposed network.