EPR-Net: Enhanced Patch Representation Network for Point Cloud Normal Estimation

doi:10.1016/j.cad.2025.103944

EPR-Net: Enhanced Patch Representation Network for Point Cloud Normal Estimation

2025-08-06

0

PRE

AI

DOI：10.1016/j.cad.2025.103944

Original

Original request for help

Save

Abstract

En 中文

• A novel GraphFormer module is designed for learning feature embeddings to effectively represent the geometric structure of local patches. The module employs the PoolFormer architecture and incorporates graph convolution with adaptive kernels, enabling the model to extract discriminative features in sharp feature regions (e.g., edges and corners). • A pyramid dynamic graph update strategy is introduced to further enhance feature integration. This strategy performs multi-scale geometric information fusion, effectively capturing both global structure and local geometry while alleviating the scale ambiguity in determining the optimal neighborhood. Meanwhile, at each network layer, this strategy expands the model’s receptive field by recomputing the k-nearest neighbors for each candidate point, enabling the model to capture potential long-range semantic characteristics in local patches. • Experiments on various datasets demonstrate the superiority of the proposed EPR-Net in point cloud normal estimation.

Keywords:

GraphFormer

feature embedding

geometric structure

pyramid dynamic graph

point cloud normal estimation

EPR-Net: Enhanced Patch Representation Network for Point Cloud Normal Estimation

Abstract

Journal

Organization