NeuLF-Net: A Neural Latent Fusion Network for 3D Surface Reconstruction

Learning-based implicit neural networks have achieved inspirational perfor-mance on point cloud surface reconstruction. To reconstruct continuous sur-faces from raw, discrete point clouds, existing methods typically project point clouds to grid latents or directly encode them as point latents. Howev-er, these methods rarely combine grid latents and point latents effectively and typically only perform simple topological transformations that ignore the spatial positional information of points, which seriously restricts the ability to capture fine details. In addition, traditional linear interpolation fails to sufficiently consider the global spatial information when inferring features of spatial points in sparse regions, resulting in a complete loss of expressiveness in some regions. In this paper, we propose a novel neural latent fusion net-work, named NeuLF-Net. The network serves as an end-to-end surface re-construction framework, efficiently retaining the spatial encoding ad-vantages of grid latents while capturing the fine-grained descriptive power of point latents. Specifically, we introduce a Neighbor Grid Enhancement Lay-er, which fully utilizes the neighbor information of the grid latents and point latents to enable enhancement of the two latents type. Furthermore, we de-sign a novel adaptive interpolation strategy that exhibits better adaptability for point cloud spatial feature extraction. We extensively evaluate our pipe-line with previous methods on three datasets including ShapeNet, Synthetic Rooms and ScanNet. Both quantitative and qualitative analyses demonstrate that NeuLF-Net substantially enhances the overall quality of point cloud re-construction. From a visual perspective, the reconstruction results appear more realistic.