U-CAN: Unsupervised Point Cloud Denoising with Consistency-Aware Noise2Noise Matching

TL;DR

U-CAN tackles unsupervised point cloud denoising by learning a multi-step denoising path for each point using Noise2Noise-style matching across multiple noisy observations. It introduces a denoising consistency constraint to stabilize local geometry and demonstrate cross-domain applicability by transferring the constraint to image denoising. The approach reports state-of-the-art performance among unsupervised methods on synthetic and real-world data and achieves competitive results with supervised methods, while also enabling unsupervised point upsampling. By leveraging Earth Mover Distance for one-to-one point correspondences, U-CAN captures fine-grained denoising patterns and aligns predictions across diverse noisy views, reducing ambiguities in surface location. The work suggests broad applicability beyond 3D and potential impact on 2D denoising tasks.

Abstract

Point clouds captured by scanning sensors are often perturbed by noise, which have a highly negative impact on downstream tasks (e.g. surface reconstruction and shape understanding). Previous works mostly focus on training neural networks with noisy-clean point cloud pairs for learning denoising priors, which requires extensively manual efforts. In this work, we introduce U-CAN, an Unsupervised framework for point cloud denoising with Consistency-Aware Noise2Noise matching. Specifically, we leverage a neural network to infer a multi-step denoising path for each point of a shape or scene with a noise to noise matching scheme. We achieve this by a novel loss which enables statistical reasoning on multiple noisy point cloud observations. We further introduce a novel constraint on the denoised geometry consistency for learning consistency-aware denoising patterns. We justify that the proposed constraint is a general term which is not limited to 3D domain and can also contribute to the area of 2D image denoising. Our evaluations under the widely used benchmarks in point cloud denoising, upsampling and image denoising show significant improvement over the state-of-the-art unsupervised methods, where U-CAN also produces comparable results with the supervised methods.

Figures (8)

• Figure 1: Overview of our method. (a) We design a multi-step denoising framework to gradually filter the noisy point cloud. (b) We introduce a novel learning schema for unsupervised learning of point cloud denoising by proposing two constraints, i.e., Noise to Noise Matching loss and Denoising Consistency loss.
• Figure 2: Illustrations on the effect of proposed constraint on denoising consistency. The noise errors indicate the Chamfer distance between the denoised and the clean point clouds.
• Figure 3: Transferring the denoising consistency constraint of U-CAN to the unsupervised image denoising.
• Figure 4: Visual comparisons under PUNet dataset. The noise errors at each point is shown in color, where the points closer to the ground truth surface are represented with bluer color, indicating lower error. And those with higher error are represented with redder color.
• Figure 4: Ablation studies on the framework and loss designs.