Unsupervised point cloud registration using heat diffusion serialization and deeply interactive attention

Point cloud registration is a fundamental problem in robotics and computer vision. Recently, Mamba, a representative State Space Models (SSMs), have emerged as an efficient point cloud modeling method due to linear complexity. To enable Mamba to process point cloud data more effectively, we propose a novel heat diffusion serialization method to convert point clouds into 1-D sequences while ensuring sequence invariance under transformations. Furthermore, to enhance the long sequence modeling capabilities of Mamba, we introduce Learnable Tokens to actively aggregate global context and filter extraneous noise during the bi-directional SSM scan. A Deeply Interactive Attention module is further designed to align and fuse local features with global context. Extensive experiments on clinical orthopedic dataset, ModelNet40, 7Scenes, and ScanObjectNN demonstrate that our method achieves state-of-the-art (SOTA) performance.