Local-Global based Deep Registration Neural Network for Rigid Alignment

Three-dimensional registration is a well-known topic in computer vision that aims to align two datasets (e.g. point clouds). Recent approaches to this problem are based on learning techniques. In this paper, we present an improved solution to the problem of registration with a novel architecture that, given two 3D point clouds as input, estimates the rotation to map one into the other. The network architecture is conceptually divided into two parts, the first part is a feature selection based on PointNet and PointNet++. The second part estimates the rotation with Euler angles by calculating the correspondences with a FlowNet-based network, and finally the rotations in yaw, pitch, and roll. The generalization capability of the proposal allows mapping two point clouds in a wide range of angles with a stable error over the whole range. Experiments have been carried out using the ModelNet10 objects dataset, varying the axis and the angle of rotation to provide a sufficiently complete evaluation of the architecture. The results show an average distance Mean Square Error of 4.94 x 10(-5) within a unit sphere and a rotation error of 1.18 degrees. The results with noisy point clouds are sufficiently accurate providing the network was trained only with noise-free data, demonstrating good generalization of the approach.