Source-linked AI summary

Score-Based Point Cloud Denoising

Shitong Luo, Wei Hu

arXiv:2107.10981v5cs.CV

TL;DR

Point clouds are noisy, which harms downstream reconstruction and analysis, while existing denoising approaches face geometric-prior and detail-preservation challenges. The paper estimates the score of a noise-convolved point distribution from noisy inputs and uses gradient ascent for denoising. Experiments report superior performance under varied noise models and potential for point cloud upsampling.

  • Problem

    Noise perturbs point clouds and affects downstream reconstruction and analysis, while existing methods face a trade-off between geometric-prior reliance, detail preservation, and denoising effectiveness.

  • Method

    The method estimates the score ∇x log[(p * n)(x)] of the noise-convolved distribution from noisy point clouds and denoises them by gradient ascent toward its mode.

  • Results

    The proposed model outperforms state-of-the-art methods under a variety of noise models and shows potential for point cloud upsampling.

  • Takeaways & Limitations

    Score-guided gradient ascent provides a denoising framework that the paper reports as robust to artifacts such as shrinkage and outliers.

  • Takeaways & Limitations

    The analysis assumes continuous noise with a unique mode at 0, although experiments report strong performance in some cases where these assumptions do not hold.

Abstract

from arXiv · show

Point clouds acquired from scanning devices are often perturbed by noise, which affects downstream tasks such as surface reconstruction and analysis. The distribution of a noisy point cloud can be viewed as the distribution of a set of noise-free samples $p(x)$ convolved with some noise model $n$, leading to $(p * n)(x)$ whose mode is the underlying clean surface. To denoise a noisy point cloud, we propose to increase the log-likelihood of each point from $p * n$ via gradient ascent -- iteratively updating each point's position. Since $p * n$ is unknown at test-time, and we only need the score (i.e., the gradient of the log-probability function) to perform gradient ascent, we propose a neural network architecture to estimate the score of $p * n$ given only noisy point clouds as input. We derive objective functions for training the network and develop a denoising algorithm leveraging on the estimated scores. Experiments demonstrate that the proposed model outperforms state-of-the-art methods under a variety of noise models, and shows the potential to be applied in other tasks such as point cloud upsampling. The code is available at \url{https://github.com/luost26/score-denoise}.

1. Introduction

Point cloud denoising is important but difficult because noise deforms structures and point clouds are irregular and unordered. The paper proposes estimating the score of a noise-convolved distribution and using gradient ascent to move noisy points toward the clean surface.

  • Motivation: Noise deforms underlying point-cloud structures, affecting downstream tasks such as rendering, reconstruction, and analysis.
  • Motivation: Point cloud denoising is challenging because point clouds are irregularly sampled from continuous surfaces and have irregular, unordered characteristics.
  • Existing approaches: Prior optimization-based methods rely heavily on geometric priors and can trade off detail preservation against denoising effectiveness.
  • Proposed approach: The proposed paradigm models noisy points with the convolution (p * n)(x), whose mode corresponds to the underlying clean surface, and denoises by gradient ascent on its log-probability.
  • Proposed approach: The method estimates the unknown score ∇x log[(p * n)(x)] from noisy point clouds using a neural network, then applies the estimated score for denoising.
  • Contributions: Extensive experiments report performance under a variety of noise models and potential application to point cloud upsampling.

2. Related Work

Prior denoising methods use geometric priors, displacement prediction, or manifold reconstruction, while score matching motivates a distinct framework that models noisy-point distributions directly.

  • Optimization-based methods rely on geometric priors and may trade detail preservation against denoising effectiveness.
  • Deep-learning denoisers commonly predict each noisy point’s displacement, but inaccurate displacement estimates can produce shrinkage and outliers.
  • Downsample-upsample reconstruction approaches can lose details, especially when noise levels are low.
  • The proposed framework instead performs gradient ascent using an estimated noisy-distribution log-density gradient, alleviating shrinkage and outliers.
  • Unlike generative score models such as ShapeGF, this method models the noise-convolved distribution and uses local score functions for arbitrary-shape generalization.

3. Method

The method estimates local scores of the noise-convolved point-cloud distribution and denoises by gradient ascent, using neighborhood-aware training and ensemble scores.

  • The noisy point-cloud distribution is modeled as p ∗ n, whose mode corresponds to the underlying clean surface.
  • A neural network estimates point-wise scores ∇x log[(p ∗ n)(xi)] from only the noisy input point cloud.
  • Local score functions focus estimation on neighborhoods around each point, supporting generalization by learning basic 3D shape fragments.
  • The score estimation unit combines each point’s feature with a nearby coordinate and predicts its localized score using an MLP.
  • Training matches predicted and ground-truth scores throughout each point’s neighborhood, so scores remain available as points move during gradient ascent.
  • An ensemble score aggregates neighboring local scores to improve robustness and reduce estimation bias before updating points by gradient ascent.
  • The step-size sequence decreases toward zero, with an initial value below 1 to support convergence and avoid over-denoising.
  • The method reports no shrinkage-induced artifacts, eliminating the need for post-processing used by some prior deep-learning denoisers.

4. Analysis

The paper models noisy point clouds as samples from a convolution of a clean-surface distribution and a noise distribution, whose mode recovers the underlying surface under mild assumptions. It then denoises by gradient ascent using an estimated score rather than the unknown density itself.

  • A noise-free point cloud is modeled as samples from a 3D distribution p supported on 2D manifolds.
  • A noisy point cloud adds noise samples to clean points, with a continuous noise density assumed to have a unique mode at 0.
  • The noisy-point distribution is the convolution of the clean distribution p and noise distribution n.
  • When the noise mode is 0 and the distribution is unimodal, the clean manifold is the mode of the noisy distribution q.
  • Denoising maximizes the log-density of each point through gradient ascent, requiring only the score ∇x log q(x).Because q is unknown at test time, the method estimates its score from noisy point clouds using a detail-preserving neural network and trains it with a score-matching objective.

5. Experiments

Experiments evaluate the method on synthetic and real-world point clouds against deep-learning and optimization-based denoisers. The method performs strongly across noise types, preserves details qualitatively, and also supports point-cloud upsampling.

  • Setup: Training uses 20 meshes sampled at 10K–50K points and perturbed with Gaussian noise at 0.5%–2% of the bounding-sphere radius.
  • Setup: Evaluation uses PU-Net and PointCleanNet test sets, multiple synthetic noise models, and the real-world Paris-rue-Madame laser-scanner dataset.
  • Setup: The method is compared with deep-learning denoisers PointCleanNet and DMRDenoise and optimization-based denoisers including bilateral filtering, jet fitting, MRPCA, and GLR.
  • Quantitative Results: Under isotropic Gaussian noise, the model significantly outperforms previous deep-learning methods in all settings and surpasses optimization-based methods in most cases.Noise standard deviations range from 1% to 3% of the shape’s bounding-sphere radius; evaluation uses Chamfer distance and point-to-mesh distance.
  • Quantitative Results: Despite Gaussian-only training, the denoiser generalizes to unseen simulated LiDAR noise and outperforms previous methods.
  • Quantitative Results: Across non-isotropic Gaussian, uni-directional, Laplace, uniform, and discrete noise, the model outperforms competing baselines in most settings.
  • Qualitative Results: Visual comparisons report cleaner results, better detail preservation, and greater outlier robustness than competing methods under Gaussian and simulated LiDAR noise.On Paris-rue-Madame, the result is cleaner and smoother than PCNet while preserving details better than DMRDenoise; the trajectory shows convergence toward the mode of p ∗n.
  • Ablation Studies: All three main design components contribute positively to denoising performance in ablation studies.The components are the score-based update, neighborhood-covering training objective, and ensemble score function.

6. Conclusions

The paper presents point cloud denoising as score-guided gradient ascent on a noise-convolved distribution. Its results also indicate potential beyond denoising, including point cloud upsampling.

  • The method models noisy point clouds as samples from a noise-convolved distribution.
  • A neural network estimates the distribution score, which guides gradient-ascent denoising.
  • The model shows potential for other point cloud tasks such as upsampling.
Loading 2107.10981v5…