PyramidPCD: A novel pyramid network for point cloud denoising

Abstract

Point cloud denoising, which aims to restore high-quality point clouds from noisy input, is an ingredient in various fields, including 3D mapping, 3D vision, and structured modeling. In this study, we present a feature pyramid network that can effectively remove noise while accurately preserving structural-to-detailed shape characteristics at varying scales. Our method is built upon the key insight that the coarser scale in the feature pyramid naturally contains more primary structures, while the finer scale provides more shape details. This discovery motivates us to progressively upgrade the current-scale feature with coarser-scale features to preserve structures while recovering details from the finer scale. To this end, we present a U-Net-shaped architecture that incorporates structure-aware and detail-preserving units at multiple scales for feature-preserving point cloud denoising. The structure-aware unit can enhance the current-scale feature by applying structural guidance from the coarse level of the feature pyramid, while the detail-preserving unit can learn a more comprehensive representation that incorporates the finer-scale feature in the pyramid. Extensive experiments conducted on publicly available benchmarks demonstrate that the proposed approach achieves state-of-the-art performances and outperforms the competing methods. Our source code and data are available at https://github.com/ForestNobear/PyramidPCD.