Pre-training meets iteration: Learning for robust 3D point cloud denoising

Abstract

Point cloud denoising is a crucial task in remote sensing and 3D computer vision, which has a significant impact on downstream tasks based on high-quality point clouds. Currently, although deep-learning-based point cloud denoising methods have demonstrated outstanding performance, their cross-dataset performance and the robustness to high-level noise remain limited. In this letter, we propose a framework called pre-training meets iteration (PMI). It presents a novel perspective that leverages point cloud pre-training for feature encoding under an iterative learning framework for point cloud denoising. Our framework exhibits robust feature encoding capabilities with pre-training. The iterative denoising architecture progressively refine data through multiple iterations to reduce noise at various levels. Under the PMI framework, we further propose a method called PMI-MAE-IT based on point masked auto-encoder and iterative neural network. The experimental results have demonstrated the outstanding cross-dataset performance of our method. Specifically, compared with state-of-the-art denoising networks, our method achieves competitive performance on the PUNet dataset, and the best performance when tested on the unseen Kinect dataset.The source code can be found at: https://github.com/hb-zhao/PMI.