Accurate and Efficient Surface Reconstruction from Point Clouds via Geometry-Aware Local Adaptation

TL;DR

The paper addresses accurate and efficient surface reconstruction from point clouds, including non-watertight geometries, by introducing curvature-aware local adaptation. It modulates patch radius, query spacing, and resampling based on local curvature, and uses unsigned distance fields with a two-stage query placement to reduce surface mixing in high-curvature areas while preserving detail in smooth regions. A key formula for local curvature is σ_n(p) = λ_0/(λ_0+λ_1+λ_2), derived from local covariance, and radius r(σ) is adjusted via a percentile-based map with predefined constants, then meshes are extracted with DCUDF. Experiments on ShapeNet Cars and DeepFashion3D Deep show improved Chamfer Distance and F1-score and reduced runtime due to fewer high-resolution queries, demonstrating robustness to complex regions while maintaining overall surface fidelity. This curvature-aware approach offers practical gains for large-scale point-cloud reconstruction pipelines by balancing detail preservation with computational efficiency.

Abstract

Point cloud surface reconstruction has improved in accuracy with advances in deep learning, enabling applications such as infrastructure inspection. Recent approaches that reconstruct from small local regions rather than entire point clouds have attracted attention for their strong generalization capability. However, prior work typically places local regions uniformly and keeps their size fixed, limiting adaptability to variations in geometric complexity. In this study, we propose a method that improves reconstruction accuracy and efficiency by adaptively modulating the spacing and size of local regions based on the curvature of the input point cloud.

Figures (4)

• Figure 1: Reconstruction pipeline. We extract a local patch assigned to each query point $q$ within a radius $r'$ based on the curvature and obtain the UDF values. Then we extract the mesh from the UDF field with DCUDFdcudf.
• Figure 2: Curvature-adaptive radius modulation. There are several phases in this strategy, separated by vertical lines. For small curvature, the support radius is dilated; it then rapidly relaxes toward the nominal radius and subsequently tapers down as curvature increases.
• Figure 3: Local query augmentation. We introduce two-stage query placement strategy based on curvature. Unevaluated query locations are filled: edge→face→cell centers.
• Figure 4: Comparison of reconstructed meshes.