FusionMatch: A Deep Fusion Method for Point Cloud Registration

Point cloud registration is a crucial task in 3-D reconstruction. Affected by noise and missing regions, it is difficult to ensure strict correspondence between points, leading to poor performance of correspondence-based methods. Global feature-based methods attempt to obtain rigid transformations through an end-to-end framework but often struggle with generalization. In this article, we propose a novel point cloud registration framework called FusionMatch, it directly generates the aligned complete point cloud by deep fusion of the input point clouds without dependence on R and t. Our approach leverages two novel components: the point cloud encoder PointPAN and the hierarchical fusion network HF. PointPAN effectively extracts the structural information of point clouds by enhancing multiscale features. HF independently fuses feature maps of different depths, thus maintaining a hierarchy of point cloud features. We conducted experiments on several datasets in terms of low overlap rate, noise, and large-angle rotation situations, and made extensive comparisons with existing methods. Experimental results show that our method achieves state-of-the-art performance at all overlap rates, even 0% overlap, and is robust to noise and rotation angles.