acceptodds
Under review as a conference paper at ICLR 2027

Coupled Subsets for Self-Supervised Point Cloud Denoising from a Single Scan

Abstract

Noisy-only learning allows point cloud denoisers to be trained without clean targets. When only a single scan of each shape is available, a natural way to obtain Noise2Noise-style supervision is to split the scan into two subsets and enforce consistency between their denoised outputs. This strategy, however, implicitly assumes that the two subsets share the same clean samples. We observe that this assumption generally fails for point clouds: subsets drawn from one scan contain different surface samples that differ in spatial coverage and local density. We term this discrepancy the coupling gap and show that a set-level consistency loss cannot be reduced below it without moving the outputs away from the clean geometry. We address the coupling gap at two levels. At the data level, a coupled paired-subset sampler builds the subsets from local point pairs, which bounds the coupling gap. At the representation level, an invertible network factorizes the latent space into a subset-invariant structure stream and a subset-specific nuisance stream, and reconstructs denoised coordinates from the structure stream alone; unlike prior factorizations, it is learned without clean targets. Training combines an Earth Mover's Distance between the denoised subsets, an entropic optimal-transport consistency between structure latents, and an input-constrained displacement regularizer. Experiments on synthetic benchmarks and real scans show that our method achieves the best overall performance among noisy-only denoisers and remains competitive with supervised denoisers. Code will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
point cloud, 3D vision

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_1fWWOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

Noise to Noise approaches for point clouds assume fully aligned matched pairs or various other strong assumptions. This paper presents a step towards making this approach more directly useful for real scans. From a single noisy scan, this approach takes a patch, and forms two subsets. At a representation level, this paper factorizes latent space into a subset invariant structure stream and subset specific nuisance stream, and reconstructs the denoised coordinates from the structure stream alone. This is novel as noise to noise on point clouds until now has made strong assumptions that mostly require the scans to come from matched pairs, or have some ground truth. Key challenge is that the two pair sets have different locations, so the loss needs to not force them together. This follows previous work on factorizing noise and structure points, but does not have ground truth points, only noise. Also does not require per shape optimization.

Claims:

  1. characterize the coupling gap between subsets of a single noisy scan and show how sampling discrepencies introduce bias into set level consistency supervision
  2. Propose couplet paired-subset sampler whose pairs bound this gap, and a structure nuisance invertable factorization that learns shared geometry from noisy observations alone.
  3. Experiments.

The paper constructs coupled subsets from the same noisy patch. As both sample same underlying surface, should share stable geometric structure. Dual stream invertible architecture to factorize latent representation into structure and nuisance code. Uses a loop of geometric consistency to ensure the denoising process goes through latent consistency. Learn a denoising operator – map noisy to denoised patch. Steps: 1) construct clean X1 and X2 with reduced discrepancy of spatial coverage and density, 2) Dual stream backbone maps into structure and nuisance representations, 3) training with composite objective: geometry consistency in Euclid space, latent structure consistency, displacement regularization. Use EMD / Wasserstein distance with squared Euclidean cost to give coupling gap D_{EMD}(X1, X2) (X*1 is unknown clean).

Prop 1 coupling gap - D of empirical denoised outputs is less than or equal to sum of distances from Y to ideal for both of them.

Algorithm for coupled pair sampler: choose a pair within a voxel; otherwise get kNN for a point, and get the nearest available point.

Then disentangle extending Mao 20222 single stream with clean targets to the dual invertible stream without clean targets here. This explicitly routes stable geometric information and input-dependent nuisance variation through separate streams (no clean targets) by reconstructing point coordinates from the structure representation.

Training objectives. Impose consistency on denoised point coordinates and latent structure representation: Geometry loss aligns denoised subsets in Euclidean space and latent structure loss aligns structure codes in latent space. Loss for geometric consistence is the square of the EMD distance between estimated pointset 1 and estimated pointset 2. Charbonnier penalty on distance between estimated points and corresponding input points . Latent structure consistency. Encourage subsets to have similar structure codes in a permutation invariant manner – uses Sinkhorn distance. So, latent must be similar for each matched point, so enforcing spatial consistency of the latent space. The distance between the denoised points is minimised, which would map points to the same space on their own, but this is counter-balanced by The Charbonnier penalty that penalizes the denoised points going to far from their original point (assuming the match is successful).

Critical strengths
  1. The paper addresses a clear issue in the noise-to-noise approach for point clouds.

  2. The approach is interesting, and addresses it in a novel and convincing way.

  3. The separating out the nuisance approach is also interesting.

  4. The results that come from this are strong.

Critical weaknesses

Proposition 1 is not very clear. It's not given a lot of context to make clear why it is important in the paper, and the explanation supporting it is also a bit unclear.

The coupled paired-subset sampler seems a reasonable way to initialise matches, but under noise it is clear that it could get an incorrect pair assignment (it finds the closest, with noise the closest may not actually be correct). It's not clear if there is a way for the matches to recover from this using Sinkhorn, or if these are always binding. The paper says Sinkhorn distance, is this simply from the pre-assigned matches or is rematching possible?

Questions for authors

Is the coupled pair-subset sampler binding for all points or can they change? How would errors from this be addressed in the architecture.

Reviewer_ecDeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a self-supervised point cloud denoising framework that learns from a single noisy scan per shape without requiring clean targets or other observations. The paper identifies a fundamental issue in subset-based consistency learning for point clouds, termed the coupling gap. Consequently, directly enforcing consistency between the denoised subset outputs may introduce geometric bias. To address this issue, the paper proposes two complementary components. First, a coupled paired-subset sampler constructs spatially balanced subsets. Second, a dual-stream invertible neural network separates latent representations into a structure stream and a nuisance stream. Experiments on synthetic benchmarks and real scans demonstrate competitive performance compared with existing noisy-only denoising methods. Ablation studies further support the effectiveness of the proposed method.

Critical strengths
  1. A well-motivated formulation of the coupling gap in single-scan self-supervised denoising. The paper identifies an important limitation of directly applying subset consistency to point clouds, termed the coupling gap. This formulation provides a useful explanation of why minimizing set-level consistency alone may introduce geometric bias. The proposed paired-subset sampler directly addresses this issue by improving the spatial correspondence between subsets. This connection between the problem formulation and the sampling strategy is a meaningful contribution.
  2. Strong empirical performance across different evaluation settings. The proposed method achieves strong results on both PU-Net and PC-Net under multiple Gaussian noise levels. Its performance also generalizes well to non-Gaussian noise distributions and real Kinect scans. The ablation studies indicate the effectiveness of the proposed components. These results provide convincing empirical evidence that the framework is effective for noisy-only point cloud denoising.
Critical weaknesses
  1. Limited methodological novelty and insufficient validation. The key components, including local point pairing, invertible latent factorization, and consistency-based self-supervision, are closely related to existing work. Although their combination is meaningful, the dual-stream INN shows only modest improvements over the single-stream variant in Table 3. Moreover, the claimed structure–nuisance disentanglement lacks direct representation-level evidence, and comparisons with recent noisy-only methods remain limited. These issues weaken the paper's methodological significance.
  2. The coupling-gap analysis does not fully justify the proposed sampling strategy under noisy observations. Section 3.2 relies on the Wasserstein triangle inequality, and the bound in Eq. (4) depends on distances between the unknown clean counterparts of paired points. However, the sampler constructs pairs from noisy coordinates. It is therefore unclear how reliably proximity in noisy observations translates into proximity on the underlying clean surface, especially under strong noise, non-uniform sampling, or sharp geometric features. While the bound itself is valid, its practical relevance to the proposed noisy-coordinate pairing strategy remains insufficiently established.
Questions for authors
  1. Given the modest improvements of the dual-stream INN over the single-stream variant in Table 3, could the authors provide direct evidence of structure–nuisance disentanglement and clarify the methodological distinction from prior invertible denoising methods?
  2. Since the bound in Eq. (4) is defined using unknown clean point positions, could the authors provide an empirical analysis of the actual coupling gap under different sampling strategies and noise levels? In particular, how does the relationship between noisy-point proximity and clean-surface proximity change as the noise level increases?
Reviewer_767qOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a method based on Noise2Noise-style supervision that uses two subsets of the same point cloud for more robustness. These subsets encode variations in noise and density. Training with loss terms for EMD, displacement, and latents enables convergence toward a mean, which is assumed to be close to the original surface. The proposed coupled paired-subset sampler enforces that both subsamples have similar statistics. This method is unsupervised and runs only on noisy observations but can compete with supervised ones. The authors did an extensive comparison with the relevant SOTA on various point cloud types with clear advantages.

Critical strengths

The comparison is solid, and the results show a clear advantage over the SOTA.

Critical weaknesses

The method, with its two subsamples linked by EMD, assumes that the noise statistics, i.e., the distribution center, are locally at least similar. This might be true in most cases, and the method overall can probably compensate a lot, but an explicit verification would still be great. Therefore, the authors should extend the (loss) ablation with a real scan (and a non-Gaussian scan if possible).

Questions for authors

Please run the loss ablation on a real scan and show the numbers. (see Critical Weaknesses above)

Reviewer_SbUoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies self-supervised point cloud denoising under a single-observation-per-shape setting. The authors identify a discrepancy between the clean point sets underlying two subsets of the same noisy scan, termed the coupling gap. They propose a spatially coupled subset sampler and a dual-stream invertible network that reconstructs denoised coordinates after suppressing the nuisance code. Training combines output-space EMD consistency, latent-space optimal-transport consistency, and an input-anchored displacement regularizer. The reported results are competitive across synthetic benchmarks, unseen noise distributions, and real scans. The sampler ablation is particularly encouraging. However, the central theoretical interpretation is stronger than what Proposition 1 establishes, and important aspects of the single-observation protocol and evaluation metrics remain unclear. The evidence for structure–nuisance disentanglement is also insufficient.

Critical strengths

S1. Relevant problem and practical inference setting.

Learning a generalizable denoiser without clean training targets is practically valuable. The proposed method uses feed-forward inference on unseen shapes and does not require per-shape test-time optimization.

S2. Reasonable and empirically supported subset construction.

Constructing spatially coupled subsets is a sensible alternative to random splitting. Table 3 reports consistent improvements over random and FPS-based splitting across the evaluated noise levels, providing useful evidence that the supervision construction matters.

S3. Broad empirical evaluation.

The experiments cover PU-Net, PC-Net, non-Gaussian corruptions, and real scans. According to the reported numbers, the method performs strongly relative to the included noisy-only baselines.

Critical weaknesses

W1. Proposition 1 establishes deviation from clean samples, not necessarily deviation from clean geometry.

The inequality in Eq. (2) is valid, but its interpretation is overstated. It shows that reducing the distance between the outputs below the coupling gap requires at least one output to differ from its corresponding clean empirical point set. It does not establish that the output must leave the underlying clean surface or degrade its geometric fidelity.For example, consider two singleton clean subsets located at (0,0,0)(0,0,0) and (2a,0,0)(2a,0,0) on a plane. Moving both outputs to (a,0,0)(a,0,0) reduces their consistency distance to zero while keeping both outputs exactly on the clean surface. Analogous examples can be constructed with larger point sets.Thus, the paper needs to distinguish recovery of the original clean samples from recovery of the underlying surface. This distinction affects the central claims in the abstract, Section 3.2, and conclusion. A correct but limited triangle-inequality argument should not be presented as a general guarantee of geometric distortion.

W2. The proposed sampler does not establish the claimed control of the clean coupling gap.

Equation (4) follows because an optimal matching cannot cost more than a specified matching. This inequality applies to any valid pairing; the substantive question is whether the proposed pairing makes its clean-space right-hand side small.However, the sampler operates on noisy coordinates. Proximity in the observed point cloud does not necessarily imply proximity between the corresponding clean samples, particularly near thin structures, nearby surfaces, or under substantial noise. The random completion stage further weakens a locality-based interpretation.The experiments show that the sampler improves denoising, but they do not directly measure the clean coupling gap, the tightness of the proposed bound, or incorrect cross-surface pairings. Consequently, the claim that performance improves specifically because the clean coupling gap is reduced remains unverified.

W3. The single-observation protocol and use of clean validation information require clarification.

Appendix A.3 states that the Gaussian noise standard deviation is sampled during training. It is unclear whether each training shape is corrupted once and then held fixed, or whether fresh corruptions are generated across epochs or patch accesses. These protocols are materially different: the latter exposes the model to multiple noisy realizations of the same underlying shape.The manuscript also selects the checkpoint with the lowest validation CD, which, under its metric definition, uses clean reference point clouds. This does not invalidate a training objective that avoids clean targets, nor does it establish test-set leakage. Nevertheless, the paper must distinguish no clean supervision in the optimization objective from no clean data in model selection. The corresponding conditions for the baselines should also be disclosed.

W4. The evidence supports useful latent truncation, but not established structure–nuisance disentanglement.

The experiments in Table 7 do not adequately demonstrate the claimed semantic factorization. Full-code reconstruction is expected from an invertible backbone, and improved reconstruction after zeroing the nuisance code is directly encouraged by the training procedure. Poor reconstruction after zeroing the structure code is also not strong evidence that the nuisance code contains no geometric information, especially given the stated allocation of coordinate channels to the structure stream.In addition, latent consistency may be affected by representation scale. If the network can rescale structure codes and compensate in the inverse mapping, the latent transport cost may decrease without a corresponding improvement in structural alignment. The manuscript does not provide enough information about latent normalization or scale control to assess this issue.More direct representation diagnostics are needed, or the disentanglement claims should be weakened.

W5. The P2M definition creates uncertainty about cross-method comparability.

Equation (16) includes both a point-to-mesh term and an additional face-wise term. The latter is not the usual one-sided point-to-surface error and weights mesh faces equally rather than by area, making its behavior potentially dependent on tessellation.This is particularly important because U-CAN results are copied from the original paper. The authors need to establish that all reported P2M values use the same definition, normalization, and implementation. If Eq. (16) is only a documentation error, it should be corrected explicitly; otherwise, the comparability of the reported rankings requires justification.

W6. The architectural contribution is not sufficiently isolated.

The method builds on invertible point-cloud restoration architectures and feature augmentation from PD-Flow/PD-LTS, but the “Single-INN” baseline is not described sufficiently to determine whether the comparison isolates the dual-stream design. Details about feature encoders, parameter counts, depth, and computational budgets are needed.Some improvements are also small. At 3% noise in Table 3, the dual-stream model improves over Single-INN from 3.842 to 3.798 in CD and from 1.862 to 1.848 in P2M. With only one reported random seed, it is difficult to judge the robustness of these differences.The patch-size analysis has an additional potential confound: the geometry loss is summed over points, whereas the displacement loss is averaged. Changing the subset cardinality can therefore change the relative loss weighting, not only the amount of geometric context.

Questions for authors

Q1. Single-observation protocol.

Is each training shape assigned exactly one fixed noisy point cloud for the entire training process? Please specify when noise is generated, whether clean points are resampled, and whether fresh corruptions are introduced across epochs or patch accesses.

Q2. Theoretical claim.

Do the authors agree that Proposition 1 establishes deviation from the original clean empirical point sets rather than necessarily from the underlying clean surface? What additional assumptions would be required to support the stronger geometric claim?

Q3. Direct coupling-gap evidence.

On synthetic data, can the authors report the actual clean coupling gap and the clean pair-distance bound for random, FPS, and coupled sampling? How frequently is random completion used, and how often do pairs connect different nearby surface structures?

Q4. Model selection.

Which clean validation data are used for checkpoint and hyperparameter selection? Are the same resources available to all baselines, and how does performance change under a selection procedure that does not use clean references?

Q5. Metric implementation.

Does the evaluation code implement Eq. (16) exactly? If so, how was compatibility with externally reported P2M values established? If not, what is the precise implemented metric?

Q6. Latent consistency and disentanglement.

What exact Sinkhorn objective, regularization parameter, and iteration count are used? Are structure codes normalized or otherwise scale-constrained? What evidence distinguishes semantic factorization from a useful but non-identifiable latent bottleneck?

1 comment

Dear Reviewer SbUo,

We sincerely appreciate the time and effort you have devoted to reviewing our paper.

As we are past the midpoint of the discussion period and approaching the deadline, we would be grateful if you could let us know whether our responses have adequately addressed your concerns or if any further clarification would be helpful.

Thank you again for your time and consideration.

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.