PointFilterNet: A Filtering Network for Point Cloud Denoising

Point clouds obtained by 3D scanning or reconstruction are usually accompanied by noise. Filtering-based point cloud denoising methods are simple and effective, but they are limited by the manually defined coefficients. Deep learning has shown excellent ability in automatically learning parameters. In this paper, a filtering network named PointFilterNet (PFN for short) is proposed to denoise point clouds by the combination of filtering and deep learning. Instead of directly outputting denoised points using networks, PFN generates the filtering denoised points through learned coefficients. Specifically, PFN outputs three coefficient vectors. These coefficient vectors are then used to filter the coordinates of points in the neighborhood of the noisy point. PFN consists of two parts: an outlier recognizer and a denoiser, both of which generate different but indispensable filtering coefficients. The outlier recognizer reduces the interference of outliers by assigning small coefficients to them. The denoiser is designed to progressively denoise point clouds which accords to the perception process of the human visual system. The experiments on synthetic and real scanned point clouds demonstrate that PFN outperforms state-of-the-art point cloud denoising works. Compared with DMR, the Chamfer distance of PFN is reduced by 22.07% on PCN-dataset.