Title: Hierarchical Space Partition for Surface Reconstruction

Abstract: Creating compact polygonal representations from point clouds remains a central challenge in computer graphics and 3D vision. However, the inherent constraints of LiDAR scanning, such as limited range and occlusions, frequently result in the loss of essential scene data, which compromises reconstruction fidelity. To mitigate this issue, we introduce a plane assembly approach designed to restore missing details without sacrificing model compactness. Our method categorizes extracted scene planes into three distinct groups: highly visible, barely visible, and invisible. The invisible planes, identified through structural scene analysis, serve to recover the absent information. These three categories establish a hierarchy of growth priorities. Planes expand based on their assigned priority, facilitating a progressive, hierarchical partitioning of the space. From this partition, we derive a watertight polygonal mesh utilizing min-cut optimization. Experimental evaluations on standard public datasets demonstrate that our method outperforms existing mainstream techniques in both effectiveness and accuracy. Further details and resources can be found at https://hsr-3dv.github.io/.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC