Boosting fine-grained feature fusion in 3D point cloud registration

Existing point cloud registration methods have achieved significant progress through transformer architecture. However, these methods often overlook the fine-grained structural information in local features, which limits their adaptability to complex scenes. To address this issue, we propose a fine-grained module that enhances the receptive field through hierarchical feature fusion. This approach provides finer-grained feature information and improves the accuracy of point cloud registration. First, a multi-scale hierarchical feature fusion module is designed to capture fine-grained feature and expand the receptive field. Second, this module is integrated into the REGTR backbone network to enhance feature correlation. Finally, an efficient and accurate registration strategy is proposed by enhancing the contribution of high-probability overlapping features. Comprehensive experiments on both indoor (3DMatch, ModelNet40) and outdoor (MCD) benchmarks demonstrate the method's effectiveness. Compared with REGTR baseline, our method achieves relative error reductions of 17.6% and 8.9% on 3DMatch and ModelNet40 respectively, while maintaining competitive computational efficiency. Consistent performance improvement on the outdoor MCD dataset further validates the method's effectiveness across diverse scenarios.