An iteration-based interactive attention network for 3D point cloud registration

Abstract

Several learning-based methods dedicated to 3D point cloud registration have been developed recently. However, the performance of such methods is limited due to a lack of effective feature interaction mechanisms. To overcome this problem, an iteration-based interactive attention network is proposed here for 3D point cloud registration that can effectively learn the overlapping features of point cloud pairs and improve the registration accuracy. Concretely, we first designed a condensed global information extractor to achieve aggregated global features. Meanwhile, to reduce the amount of calculation required for subsequent modules while maintaining high-level characteristic information, the number of point cloud features is reduced. Furthermore, an overlapping feature iterative attention module and a discriminant feature iterative expansion module are designed. The main function of both modules is to implement feature interaction in the form of cross-attention and self-attention as well as to deepen interactions iteratively. The former captures overlapping information by iteratively interacting with the features of the condensed global features of one point cloud and the local geometric features of another point cloud. The latter can expand the number of features back to the initial number of point clouds through the interaction of overlapping features and local geometric features. Finally, matching features with high discriminative ability are obtained and used to formulate a soft correspondence matrix. Furthermore, weighted singular-value decomposition is adopted to obtain the rigid transformation matrix. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance.