Pose Estimation by Key Points Registration in Point Cloud

In 6D pose estimation, both template based or learning based models need template/training data corresponding to different poses. We propose a new model based on point clouds. It needs less training data and therefore is lighter, faster and precise enough for texture-less objects. We first extract key points automatically from the point cloud of the object and then generate rigid transformation invariant point-wise features of the cloud as input feature. Then we use a hierarchical neural network architecture to predict the key points coordinates corresponding to the reference pose. At last we can calculate the relative transformation between the current and the reference poses. The hierarchical structure takes into account the symmetry or invariance problem of certain object geometries.