Geometric Encoding-Based Attention Mechanism for Point Cloud Registration Network

Abstract

We propose a novel point cloud registration network called GEANet, which overcomes the issues of disregarding point cloud geometry information and inadequate utilization of geometric information by utilizing an attention mechanism-based approach and point cloud geometry encoding. Our approach starts by extracting point cloud features using Graph Neural Network (GNN) and feeding them into a point cloud geometry encoder to obtain geometric encoding. The encoding is then jointly input into the attention mechanism for feature interaction. Next, virtual point pairs and weight values are obtained by jointly calculating the point cloud features and point cloud spatial information features, such as Euclidean distance and direction vectors, and the required rigid transformation is solved through SVD. Our experimental results on the ModelNet40 dataset, including unseen point clouds, unseen point cloud categories, and Gaussian noise, demonstrate that the MSE for rotation matrices were reduced to 1.97, 1.68, and 3.70, and for translation to 0.015, 0.013, and 0.018. The findings suggest that our GEANet approach achieves higher accuracy and greater robustness than point cloud registration networks that solely rely on Transformer and do not utilize point cloud geometry encoding.