Point cloud denoising with adaptive graph convolution

Abstract

Point cloud denoising is a key aspect of current point cloud data processing. Existing deep learning-based denoising methods rely on local feature information obtained from noisy point clouds as the basis for point displacement. However, in the presence of noise, the local features of points usually cannot accurately represent the underlying surface of the point cloud. We propose a new deep learning-based denoising method using the graph convolution operator, which can compute feature aggregation weights adaptively based on feature vectors to achieve hierarchical distinction between noise and the underlying surface. Meanwhile, in order to alleviate the phenomenon that the predicted gradient fluctuates more drastically during the iterative process, we introduce the momentum rise method in the gradient iteration process to improve the stability of the solution. We evaluated our approach on both toy datasets and real world datasets. Under different types and levels of noise disturbances, our method has better results compared to other models.