Pose manifolds for efficient visual servoing

In order to adequately accomplish vision-based manipulation tasks, robotic platforms require an accurate estimation of the 3D pose of the target, which is efficiently approached by imaging techniques excessively utilizing large databases that consist of images of several objects captured under varying viewpoints. However, such approaches are characterized by large computational burden and complexity accompanied by limited capacities to interpolate between two known instances of an object. To address these issues we propose a robust 3D object pose estimation technique that entails a manifold modeling procedure based on appearance, geometrical and shape attributes of objects. We utilize a bunch-based method that is followed by a shape descriptor module, in order to establish low dimensional pose manifolds capable of distinguishing similar poses of different objects into the corresponding classes. Finally, an accurate estimation of the 3D pose of a target is provided by a neural network-based solution that encompasses a novel input-output space targeting method. We have comparatively studied the performance of our method against other related works, whilst experimental results justify our theoretical claims and provide evidence of low generalization error.