Efficient pose estimation of rotationally symmetric objects

In this paper we propose algorithms for 3D object recognition from 3D point clouds of rotationally symmetric objects. We base our work in a recent method that represents objects using a hash table of shape features, which allows to match efficiently features that vote for object pose hypotheses. In the case of symmetric objects, the rotation angle about the axis of symmetry does not provide any information, so the hash table contains redundant information. We propose a way to remove redundant features by adding a weight factor for each set of symmetric features. The removal procedure leads to significant computational savings both in storage and time while keeping the recognition performance. We analyze the theoretical storage gains and compare them against the practical ones. We also compare the execution time gains in feature matching and pose clustering. The experiments show storage gains up to 100× and execution time savings up to 3500× with respect to state-of-the-art methods.