GBPA-Net: Geometric Back-projection Probsparse Attention network for Point Cloud Registration

To address the problems of slow speed and poor robustness of traditional point cloud registration algorithms and the inability to achieve precise radiotherapy for cancer patients before and after radiotherapy, a new geometric inverse projection sparse attention network is proposed for point cloud registration. The method transforms the feature-rich and detailed GBNet feature extraction module into a feature-rich DGCNN feature extraction module, which can simultaneously extract comprehensive point cloud features and reduce the redundancy of features. In addition, sparse attention for long sequences is introduced, thus reducing the computational complexity of traditional attention. We have conducted extensive experiments on the modelnet40 dataset to validate the effectiveness of GBPA-Net. Numerous experiments show that the method has better performance compared with other methods and demonstrate the effectiveness and robustness of the method.