Self-Geometric priors and rough-to-fine cascade optimization based on SSM for surface reconstruction from sparse point cloud

In the past few decades, we have greatly improved our ability to process point cloud data with the development of deep learning. Surface reconstruction has made significant progress, but there are still some challenges, such as meshes generated with holes, sensitivity to noise, and many approaches being hard to produce desirable results when faced with a wide variety of unseen classes. We designed the framework to be able to generate watertight meshes from the point cloud based on self-geometric priors and Mamba from rough to fine. First, we utilize a module based on adaptive graph convolution to upsample the point cloud for enhanced geometric information, subsequently, we employ a straightforward algorithm to swiftly generate a rough initial mesh. Following this, the self-geometric priors deform module, utilizing Mamba, is trained to iteratively adjust the vertex positions of the initial mesh. Eventually, a high-precision watertight mesh that conforms to the geometry of the point cloud is generated. We conducted a large number of experiments, and finally proved the effectiveness of each module of our framework, and improved the accuracy of point cloud reconstruction compared with existing methods.