Noise2Score3D: Tweedie's Approach for Unsupervised Point Cloud Denoising¶
Conference: ICCV 2025
arXiv: 2503.09283
Code: GitHub
Area: 3D Vision
Keywords: Point cloud denoising, unsupervised learning, Tweedie's formula, score function, total variation
TL;DR¶
This paper proposes Noise2Score3D, a fully unsupervised point cloud denoising framework based on Tweedie's formula. It learns the score function directly from noisy data and achieves single-step denoising, while introducing point cloud total variation to estimate unknown noise parameters.
Background & Motivation¶
3D point clouds are frequently corrupted by noise due to sensor errors and environmental factors. Existing deep learning-based denoising methods face two fundamental challenges:
Supervised methods depend on paired data — obtaining clean data in real-world scenarios is difficult or impossible.
Existing unsupervised methods suffer from significant limitations: - TotalDn relies on spatial priors and involves slow iterative processes. - The "score" in ScoreDenoise is based on displacement rather than likelihood, precluding the use of Bayesian statistical tools. - Poor generalization (retraining required across datasets and noise levels). - Low inference efficiency (iterative denoising required).
Noise2Score has demonstrated the power of Tweedie's formula in 2D image denoising: by reformulating denoising as a score function estimation problem, it offers a theoretically elegant and flexible framework. This paper extends this idea to 3D point clouds.
Method¶
Theoretical Foundation — Tweedie's Formula¶
For Gaussian noise: \(\mathbf{y} = \mathbf{x} + \sigma\epsilon\)
Tweedie's formula yields an explicit expression for the posterior expectation: $\(E_{p(\mathbf{x}|\mathbf{y})}(\mathbf{x}) = \mathbf{y} + \sigma^2 \nabla_\mathbf{y} \log p(\mathbf{y})\)$
where \(\nabla_\mathbf{y} \log p(\mathbf{y})\) is the score function of the noisy data distribution. Key advantage: given the score function and noise parameters, denoising is accomplished in a single step.
Score Estimation Network¶
Based on the KPConv architecture: - 5 encoding stages (14 blocks): 3→256→512→1024→2048 channels - 4 decoding blocks: 3072→128 - Final fully connected layer outputting a 3D score vector - 24.3M parameters
The Amortized Residual DAE (AR-DAE) loss is employed: $\(\mathcal{L}_{AR-DAE} = \frac{1}{N}\sum_{i=1}^N \|\sigma_t \cdot S(y'_i) + u\|^2\)$
where \(y'_i = y_i + u \cdot \sigma_t\) is the perturbed version and \(u \sim \mathcal{N}(0, I)\).
Unknown Noise Parameter Estimation — Point Cloud Total Variation¶
The optimal noise parameter is estimated by minimizing the total variation of the denoised result: $\(\sigma^* = \arg\min_\sigma TV_{PC}(\hat{x}(\sigma))\)$
Denoising Pipeline¶
- Training: Learn the score function \(S(y)\) from noisy point clouds.
- Denoising: \(\hat{x} = y + \sigma^2 S(y)\) (completed in a single step).
- Unknown noise: Search for the optimal \(\sigma\) over a set of candidates.
Key Experimental Results¶
Denoising Performance on ModelNet-40 (CD×\(10^4\))¶
| Method | 10K pts 1% noise CD | 10K pts 3% noise CD | 50K pts 1% noise CD |
|---|---|---|---|
| Bilateral | 5.865 | 31.034 | 3.711 |
| GLR | 6.592 | 12.890 | 1.860 |
| TotalDn | 8.079 | 29.617 | 5.044 |
| Score-U | 5.514 | 18.239 | 2.696 |
| Noise2Score3D | 4.891 | 12.456 | 1.654 |
Generalization Evaluation¶
| Training Set | Test Set | CD↓ | P2M↓ |
|---|---|---|---|
| ModelNet-40 | ModelNet-40 | Best | Best |
| ModelNet-40 | PU-Net | Still superior | Still superior |
Key Findings¶
- Achieves state-of-the-art performance among unsupervised methods, approaching supervised methods in certain settings.
- Strong generalization — a single pretrained model transfers across datasets and noise levels without retraining.
- Efficient inference — single-step denoising when noise parameters are known, outperforming iterative approaches.
- Point cloud total variation serves as an effective no-reference quality metric for automatic noise parameter selection.
Highlights & Insights¶
- Integration of Bayesian statistical tools — Tweedie's formula decouples denoising from score estimation, offering theoretical elegance.
- Single-step denoising — distinguishes this work from all existing unsupervised methods that rely on iterative procedures.
- Noise model agnosticism — the same loss function and pretrained weights apply across different noise models.
- Point cloud total variation — introduces a no-reference quality metric to point cloud denoising for the first time.
Limitations & Future Work¶
- Validation is primarily conducted under Gaussian noise; performance on other noise distributions remains to be verified.
- The KPConv-based network has a relatively large parameter count (24.3M).
- When noise parameters are unknown, a search over a candidate range is required.
Related Work & Insights¶
- Traditional methods: Bilateral filtering, low-rank approximation, graph Laplacian regularization.
- Supervised methods: PointCleanNet, ScoreDenoise, IterativePFN.
- Unsupervised methods: TotalDn, DMR-U, Score-U.
- Score matching: Denoising Score Matching, AR-DAE.
Rating¶
- Novelty: ⭐⭐⭐⭐ (Extension of Tweedie's formula to 3D point clouds)
- Technical Depth: ⭐⭐⭐⭐⭐ (Solid theoretical foundation, complete Bayesian framework)
- Experimental Thoroughness: ⭐⭐⭐⭐ (Multiple noise levels, cross-dataset generalization)
- Value: ⭐⭐⭐⭐⭐ (No clean data required, single-step denoising, automatic noise estimation)