Robust Point Cloud Normal Estimation via Multi-Level Critical Point Aggregation

In this paper, we propose a multi-level critical point aggregation architecture for 3D point cloud normal estimation. It efficiently focuses on locally important points during feature extraction by employing our Local Feature Aggregation (LFA) and Global Feature Refinement (GFR) modules. These modules can accurately identify critical surface-fitting points across local and global levels. Specifically, the proposed LFA module aims to capture local geometric information from nearby points with strong correlation in low-level features, while our GFR module explores global geometric relationships in high-level features space, focusing on critical global points. Furthermore, utilizing a stacked LFA structure, we address indistinguishable features across multiple levels, enabling deep feature aggregation. By integrating multi-level features through the GFR module, our method effectively integrates robust local geometric information into comprehensive global features. This ensures stability and accuracy in subsequent surface fitting and normal estimation tasks, even in the presence of sharp features, high noise, or anisotropic structures. Experimental results demonstrate that our method is competitive and achieves stable performance on both synthetic and real-world datasets. Our implementation is available at https://github.com/CharlesLee96/NormalEstimation.