Improvement of normal estimation for point clouds via simplifying surface fitting.

During the last decade, stimulated by the roaring success of neural networks in the realm of point clouds, learning-based normal estimation has been a dominating direction and it has attained outstanding performance. The introduction of modern neural networks significantly enhances the robustness of algorithms to noise levels and sampling scales. A method that combines least-squares surface fitting with weight-learning neural networks is particularly prominent. Despite persistent efforts, there is room for improvement. We observed that a simplified surface-fitting process can significantly improve the accuracy of this type of hybrid approach. In this study, two simple yet effective strategies are proposed to boost the accuracy of normal estimation. Specifically, a dynamic Top-K selection strategy is presented to focus on the critical points of a given patch. The selected point hammer fits a local surface using a simple tangent plane, which improves the normal estimation accuracy when encountering patches with sharp corners or complex patterns. Moreover, to further improve the quality of the estimation, a point-update strategy was delicately designed to smooth the sharp boundaries of the patches before the local surface fitting process. Thus, our method can reduce the fitting distortion while simplifying the surface-fitting process. Extensive experiments and ablation studies demonstrated the superiority of our proposed method over contemporary state-of-the-art approaches.