DFNet: Dual Feature Network for Point Cloud Registration

In point cloud registration, global and local features collectively constitute a comprehensive description of point cloud data. Previous methods for registration mostly relied on a single kind of features, and the correspondence matching based methods require inputs with distinctive local structures, so the quality of feature extraction directly impacts the accuracy of the registration results. The global feature-based methods can obtain global geometric information of the inputs, but the lack of local information can easily overlook the negative effect of non-overlapping points. In this paper, we present DFNet, an iterative network based on dual feature extraction for partial-to-partial point cloud registration. We use a dual branch structure to individually extract both global and local features, fully utilizing the geometric characteristics of the point cloud. Experiments indicate that our method outperforms other registration approaches in terms of accuracy and robustness.