Multi-scale hash encoding based neural geometry representation

Recently, neural implicit function-based representation has attracted more and more attention, and has been widely used to represent surfaces together with differentiable neural networks. However, existing neural geometry representations still suffer from slow computation speed and insufficient reconstruction accuracy when applied to surface reconstructions from point clouds and multi-view images. To alleviate these issues, we propose a multi-scale hash encoding-based neural geometry representation to effectively and efficiently optimize the surface represented as a signed distance field. To this end, a novel neural network structure is proposed by carefully combining low-frequency Fourier position encoding with multi-scale hash encoding. Accordingly, the initialization of the geometry network and geometry features of the rendering module is redesigned. Extensive experiments demonstrate that our proposed representation achieves at least 10 times speedup on the task of million-level point cloud reconstruction, and significantly improves the efficiency and accuracy on the multi-view reconstruction task. Our code and models will be available at https://github.com/Dengzhi-USTC/ Neural-Geometry-Reconstruction .