MuSic-UDF: Learning Multi-Scale dynamic grid representation for high-fidelity surface reconstruction from point clouds

Surface reconstruction for point clouds is a central task in 3D modeling. Recently, the attractive approaches solve this problem by learning neural implicit representations, e.g., unsigned distance functions (UDFs), from point clouds, which have achieved good performance. However, the existing UDF-based methods still struggle to recover the local geometrical details. One of the difficulties arises from the used inflexible representations, which is hard to capture the local high-fidelity geometry details. In this paper, we propose a novel neural implicit representation, named MuSic-UDF, which leverages Multi-Scale dynamic grids for high-fidelity and flexible surface reconstruction from raw point clouds with arbitrary typologies. Specifically, we initialize a hierarchical voxel grid where each grid point stores a learnable 3D coordinate. Then, we optimize these grids such that different levels of geometry structures can be captured adaptively. To further explore the geometry details, we introduce a frequency encoding strategy to hierarchically encode these coordinates. MuSic-UDF does not require any supervisions like ground truth distance values or point normals. We conduct comprehensive experiments under widely-used benchmarks, where the results demonstrate the superior performance of our proposed method compared to the state-of-the-art methods. [Display omitted] •Propose a novel dynamic grid representation to learn UDFs from raw point clouds.•Introduce a hybrid encoding strategy to hierarchically encode the dynamic grids.•Introduce appropriate metrics to construct loss functions.•Conduct comprehensive evaluations and achieved state-of-the-art results.