Title: SBP-Net: Learning Thin Structure Reconstruction with Sliding-Box Projections

Abstract:

The reconstruction of slender 3D architectures presents significant difficulties, primarily driven by their sparse nature, varying scales, and intricate geometries. These types of structures are prevalent across numerous fields, such as the medical imaging of vascular networks and the modeling of industrial piping systems. Although contemporary neural approaches excel at capturing dense surfaces, they frequently struggle to accurately recover fine, thin geometries. To address this, we introduce a novel reconstruction framework centered on local depth projections, which serve as an efficient and highly informative two-dimensional representation of thin structures. Our method involves moving a sliding box across the 3D model to generate local orthographic depth projections. These projections are then fed into a neural network to reconstruct the missing thin components in 2D space. Finally, these localized reconstructions are integrated back into the 3D model, yielding a cohesive and highly detailed shape. We validated our approach through experiments on pulmonary artery reconstruction from CT volumes and industrial pipeline recovery using both synthetic and real-world scans. The results indicate that our method significantly outperforms existing techniques in preserving fine structural details.

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