Learnable Point Cloud Sampling Considering Seed Point for Neural Surface Reconstruction

Reconstruction of surfaces from point clouds is essential in numerous practical applications. An approach in which neural fields are trained as surface representations from point clouds has garnered significant interest in recent years. However, these techniques present scalability issues to large scenes since they are limited in the size of point clouds that can be processed. This work proposes a learnable point cloud sampling designed to address the scalability issues. We introduce a sampling network that considers a seed point acting as the origin to sample points from a part of the scene. The seed point is one of the input points that is selected in a spatially uniform manner. This prompts a surface reconstruction network to learn the detailed geometry on partial regions of the scene. We also propose a training pipeline based on point cloud splitting and merging to avoid an increase in the memory footprint. We jointly train the sampling network and surface reconstruction network using a task loss to optimize the sampling network for the surface reconstruction task. Experimental results on scene-level datasets captured from real-world environments demonstrate that our method performs better than state-of-the-art methods in surface reconstruction.