Noise2Score3D: Tweedie's Approach for Unsupervised Point Cloud Denoising

TL;DR

Noise2Score3D introduces a fully unsupervised point cloud denoising framework that learns the score function of the noisy data distribution and applies Tweedie’s formula for one-step denoising. Built on a KPConv-based encoder–decoder, it trains with an AR-DAE loss to estimate per-point scores without clean targets, enabling robust denoising across noise levels. A key novelty is the TV_PC metric, which both guides unknown-noise-parameter estimation and provides a principled criterion for selecting the denoised output. Empirically, Noise2Score3D achieves state-of-the-art performance among unsupervised methods on ModelNet-40 and PU-Net benchmarks, generalizes to unseen datasets and LiDAR-like noise, and runs significantly faster at inference than competing unsupervised approaches, making it practical for real-world deployments.

Abstract

Building on recent advances in Bayesian statistics and image denoising, we propose Noise2Score3D, a fully unsupervised framework for point cloud denoising. Noise2Score3D learns the score function of the underlying point cloud distribution directly from noisy data, eliminating the need for clean data during training. Using Tweedie's formula, our method performs denoising in a single step, avoiding the iterative processes used in existing unsupervised methods, thus improving both accuracy and efficiency. Additionally, we introduce Total Variation for Point Clouds as a denoising quality metric, which allows for the estimation of unknown noise parameters. Experimental results demonstrate that Noise2Score3D achieves state-of-the-art performance on standard benchmarks among unsupervised learning methods in Chamfer distance and point-to-mesh metrics. Noise2Score3D also demonstrates strong generalization ability beyond training datasets. Our method, by addressing the generalization issue and challenge of the absence of clean data in learning-based methods, paves the way for learning-based point cloud denoising methods in real-world applications.

Figures (13)

• Figure 1: Training workflow of Noise2Score3D with feature extraction and score prediction by an encoder-decoder network with the AR-DAE loss. Denoising is done with Tweedie's formula using estimated scores to restore the position of point clouds (not shown here).
• Figure 2: Visualization results by different algorithms on synthetic datasets: (a) ModelNet-40 with 50k points; (b) PU-Net with 10k points; (c) PU-Net with 50k points; (d) ModelNet-40 with simulated LiDAR noise. Note that shown results with the noise scale set to 2% of the bounding sphere’s radius for the Gaussian noise and 1.5% for the simulated LiDAR noise. Points with smaller error are colored more blue, and otherwise colored yellow.
• Figure 3: Visualization results on the real-world dataset Paris-rue-Madameserna2014paris.
• Figure 4: Change of average $TV_{PC}$, $CD$ and $P2M$ values with noise parameter $\sigma$ on the PU-Net datasetyu2018PUNet .
• Figure 5: Additional visualization results of different algorithms on ModelNet-40 dataset with Gaussian noise. The noise level is set to 1%.