Scholarlink

<p>Implicit surface reconstruction from point clouds is typically achieved through signed distance function (SDF) estimation, where the surface is represented as the zero-level set of an implicit function and fitted using 3D points with supervised values near the surface. To enhance reconstruction quality, this paper improves the Neural Implicit Moving Least Squares (Neural-IMLS) framework by introducing an adaptive radius adjustment mechanism based on local point cloud density. Specifically, a k-nearest neighbor (KNN)-based density perception guides the dynamic adjustment of the ball query radius, enabling more reasonable sampling for the implicit neural network in SDF estimation. This adaptive strategy improves the recovery of complex local details and increases reconstruction accuracy. Experimental validation is conducted on Stanford 3D point cloud models, comparing the proposed method with Sign Agnostic Learning (SAL) and Neural-IMLS. Results show significant improvements: for the Bunny model, Chamfer Distance (CD) and Hausdorff Distance (HD) decrease by 27.3% and 53.1%, respectively; for the Armadillo model, CD decreases by 16.6% and HD by 55.3%. Qualitative comparisons with ground truth(GT) meshes further confirm that the proposed method achieves superior reconstruction performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Research on implicit surface reconstruction based on adaptive query radius adjustment mechanism

  • 期刊论文

  • 发表日期：

    2025年10月28日

查看全文
  • Yun-tao Zhao,
  • Xuan-zhao Yu,
  • Xin-fei Gong,
  • Yuan-qing Li

摘要

Implicit surface reconstruction from point clouds is typically achieved through signed distance function (SDF) estimation, where the surface is represented as the zero-level set of an implicit function and fitted using 3D points with supervised values near the surface. To enhance reconstruction quality, this paper improves the Neural Implicit Moving Least Squares (Neural-IMLS) framework by introducing an adaptive radius adjustment mechanism based on local point cloud density. Specifically, a k-nearest neighbor (KNN)-based density perception guides the dynamic adjustment of the ball query radius, enabling more reasonable sampling for the implicit neural network in SDF estimation. This adaptive strategy improves the recovery of complex local details and increases reconstruction accuracy. Experimental validation is conducted on Stanford 3D point cloud models, comparing the proposed method with Sign Agnostic Learning (SAL) and Neural-IMLS. Results show significant improvements: for the Bunny model, Chamfer Distance (CD) and Hausdorff Distance (HD) decrease by 27.3% and 53.1%, respectively; for the Armadillo model, CD decreases by 16.6% and HD by 55.3%. Qualitative comparisons with ground truth(GT) meshes further confirm that the proposed method achieves superior reconstruction performance.