DFPINet: Dual Feature and Progressive Interaction Network for Point Cloud Registration

Abstract

In point cloud registration, the combination of global and local features provides a comprehensive representation of point cloud data. Previous registration methods predominantly relied on a single type of feature. For instance, correspondence matching-based methods require stringent criteria for the uniqueness of input point clouds, and the quality of feature extraction significantly influences the accuracy of registration results. Global feature-based methods capture the overall geometric information of the inputs; however, the absence of local information often leads to the neglect of the adverse effects caused by non-overlapping points. In this paper, we present DFPINet, an iterative network based on dual feature extraction and progressive feature interaction for partial-to-partial point cloud registration of small-scale object point clouds. We use a dual branch structure to individually extract both global and local features, fully utilizing the geometric characteristics of the point cloud. Additionally, the progressive feature interaction operation enhances feature connections and mitigates the effect of non-overlapping points. Experimental results demonstrate that our method surpasses existing registration approaches in both accuracy and robustness.