Frequency-Domain Guided Point Cloud Normal Estimation Method

To address the insufficient global information perception in feature extraction of point cloud normal estimation under the deep learning paradigm, this paper proposes a Fourier spectrum analysis guided method for point cloud normal estimation. Specifically, local point cloud patches are first mapped from the spatial domain to the frequency domain using Fast Fourier Transform. Then, by analyzing the contribution of different frequency components, the spectrum is proportionally divided into high frequency, mid frequency, and low frequency parts. Leveraging the global structural characteristics inherent in the frequency domain enhances the accuracy of normal estimation. Finally, a learning based adaptive strategy is introduced to dynamically refine the multi frequency information, further improving the precision and robustness of normal estimation. Experimental results on error evaluation and 3D reconstruction demonstrate that compared with spatial domain information, the global properties of the frequency domain provide richer intrinsic features, and the influence of different frequency components on normal estimation varies significantly. The proposed method achieves favorable performance on three public benchmark datasets, namely PCPNet, FamousShape and SceneNN, with an average RMSE improvement of about 0.5%.