Registration Loss Learning for Deep Probabilistic Point Set Registration
Abstract: In recent years, probabilistic methods for point set registration have demonstrated promising performance. These methods represent the scene as a mixture of Gaussians that model the point set densities. The parameters of the model and registration transformations are then inferred jointly. This formulation has interesting theoretical properties, such as linear complexity in the number of used points, and it easily generalizes to joint registration of multiple point sets. In this work, we extend the probabilistic paradigm to benefit from learned features, by adding a von Mises-Fisher feature model in each mixture component. We also propose a learning strategy that directly uses the registration error as a loss, by back-propagating through the registration iterations. This is possible as the probabilistic registration is fully differentiable, and the result is a learning framework that is truly end-to-end. We perform extensive experiments on the 3DMatch and Kitti datasets. The experiments demonstrate that our approach benefits significantly from the integration of the learned features and our learning strategy, outperforming several state-of-the-art methods.
Loading