acceptodds
Withdrawn from ICLR 2027

RODR: Riemannian Orthogonal Decoupling Representation for Feature-Preserving Continuous Surface Recovery

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Continuous surface recovery from discrete, noisy point clouds is fundamental to 3D representation learning. Despite empirical progress in flow matching, score-based diffusion, and coordinate regression, existing paradigms suffer from a structural objective-geometry conflict: isotropic Euclidean losses couple extrinsic surface-normal attraction with intrinsic tangential spacing, precipitating contractive collapse and quadratic off-manifold drift. In this work, we propose Riemannian Orthogonal Decoupling Representation (RODR), a lightweight, architecture-agnostic geometric optimization framework. RODR projects spatial vector updates strictly onto local normal bundles and tangent bundles, mathematically eliminating first-order cross-subspace gradient interference. Coupled with a multi-scale geometric planarity gate, RODR dynamically throttles tangential velocity near structural discontinuities, locking in sharp CAD boundaries without over-smoothing. Crucially, continuous flow optimization under RODR enables neural velocity fields to internalize the underlying Riemannian metric directly into network parameters, avoiding the finite-sample covariance noise of test-time discrete geometric projections. Extensive evaluations across standard benchmarks, high-density CAD models (up to 200K points), deterministic regressors, and stochastic differential equations demonstrate that RODR cuts surface manifold thickness (Local Manifold Fitting Residual, LMFR) by up to 38.2%, delivering metrology-grade downstream mesh reconstructions while accelerating inference throughput with a compact 0.53M backbone.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Geometric Deep Learning, Normal Bundle Decoupling, Continuous Surface Reconstruction, CAD feature preservation, Riemannian Manifold

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_qb8kOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Riemannian Orthogonal Decoupling Representation (RODR), a geometry-aware framework for continuous 3D surface recovery from noisy point clouds. The core idea is to decouple ambient Euclidean spatial updates into normal-direction surface fitting and tangent-direction point redistribution via orthogonal subspace projections. Additionally, a multi-scale planarity gate is introduced to dynamically suppress tangential flow near sharp geometric discontinuities, preventing feature over-smoothing. The paper provides theoretical analyses of the projected dynamics and validates the method across standard denoising benchmarks, dense CAD models, and real-world LiDAR data. The evaluation reports standard reconstruction metrics as well as LMFR, a metric used to measure local surface thinness.

Critical strengths

The paper provides a clear geometric interpretation of normal-direction surface attraction and tangent-direction point redistribution. Their explicit separation into complementary subspaces offers a simple formulation for controlling off-surface deviations in dense CAD recovery.

The dense-CAD experiments show improvements in LMFR, suggesting a potential benefit in reducing artificial manifold thickness. The cross-sectional visualizations provide complementary qualitative evidence.

Critical weaknesses
  • The methodological novelty of RODR is unclear. RODR maps the fitting and uniformity forces onto normal and tangent subspaces, respectively, but the underlying normal/tangent decomposition and orthogonal projection are standard geometry. It is unclear what fundamentally new methodological ingredient RODR introduces beyond this formulation.

  • The specific meaning of “Riemannian” in this paper could be clarified. While the paper motivates the formulation through the Riemannian manifold hypothesis, the proposed decoupling itself appears to be based on the normal/tangent decomposition and orthogonal projections. It would be helpful to clarify whether “Riemannian” refers primarily to this manifold interpretation or to an additional methodological property of the proposed method.

  • In Table 2, applying the orthogonal loss alone improves LMFR but degrades the P2F metric compared to the baseline. The significant P2F gain only appears after adding the Distance Magnitude Head. A deeper analysis is needed to isolate whether the orthogonal decoupling itself is inherently effective or if it relies heavily on the distance head to prevent overshooting.

  • In Table 5 (PUNet/PCNet), the proposed method does not show a consistent advantage across all metrics/resolutions compared to recent methods. Additionally, while Appendix F explores extreme noise on a thin plate, evaluation under higher noise levels (e.g., 3%) on complex 3D CAD benchmarks is needed, as local normal/tangent subspace estimation naturally degrades under strong noise.

Questions for authors
  • What is the key methodological novelty of RODR beyond the standard normal/tangent decomposition and orthogonal projection? Could the authors clarify which component of the proposed formulation is fundamentally new?

  • What specifically does “Riemannian” refer to in RODR beyond the underlying Riemannian manifold hypothesis?

  • Could the authors clarify why applying the orthogonal loss alone degrades the P2F metric compared to the baseline, and whether the performance gain relies primarily on the Distance Magnitude Head?

  • Given the mixed results in Table 5, could the authors clarify RODR's clear operational advantage over recent methods and provide evaluation results on complex 3D CAD models under higher noise levels (3%)?

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

This paper proposes RODR, a geometry-aware framework for point-cloud surface recovery. It decouples surface fitting and point redistribution into orthogonal normal and tangent subspaces, and further uses a multi-scale planarity gate to preserve sharp geometric structures. The method is applicable to flow matching, coordinate regression, and score-based diffusion, and demonstrates strong improvements in manifold thinness, dense CAD recovery, and inference efficiency.

Critical strengths

The distinction between normal-direction surface attraction and tangential point redistribution is intuitive and physically meaningful, and the proposed orthogonal decomposition offers a principled way to separate these two update components. The ablation results also suggest that this design is beneficial for reducing manifold thickness.

The proposed geometric decoupling can be incorporated into continuous flow matching, deterministic coordinate regression, and score-based diffusion without major architectural redesign. Although the gains are not uniform across all metrics, the experiments indicate that the formulation is not restricted to a single denoising framework.

RODR achieves notable reductions in manifold thickness on dense CAD point clouds and maintains favorable computational efficiency with a relatively compact backbone. These results suggest that the method is suitable for scenarios where surface thinness and high-density processing are important.

Critical weaknesses

The paper states that RODR enables the continuous velocity field to internalize the underlying Riemannian geometry, thereby avoiding explicit geometric projection at test time. However, the current experiments mainly demonstrate that RODR performs well without test-time projection, rather than directly showing that the learned field encodes local normal/tangent geometry or an equivalent Riemannian structure.

A key motivation of RODR is that conventional Euclidean objectives introduce cross-subspace interference between surface fitting and point uniformization. Although the paper theoretically establishes the orthogonality of the projected normal and tangential updates, it does not quantify how much RODR reduces such interference.

The full RODR model achieves better performance on sharp creases and boundary regions, but the region-wise analysis compares the full method against the baseline rather than directly comparing variants with and without the planarity gate. Therefore, it remains unclear to what extent the observed sharp-feature preservation can be specifically attributed to the gate rather than to other components of RODR.

RODR performs particularly well on manifold thinness and high-density CAD recovery, but it is not consistently superior across all standard benchmarks and metrics; in some plug-and-play settings, improvements in LMFR are accompanied by slight degradation in other geometric metrics. In addition, the real-world evaluation is limited to a single mobile LiDAR dataset and lacks direct comparisons with strong denoising baselines, which weakens broader claims of state-of-the-art and real-world generalization.

Questions for authors

Could the authors provide more direct evidence supporting the claim that the Riemannian geometry is internalized into the learned velocity field?

Could the authors directly quantify the Euclidean gradient entanglement that motivates RODR?

Could the authors further isolate the effect of the multi-scale planarity gate on sharp-feature preservation?

If the Riemannian geometry is claimed to be internalized into the network parameters, why does the inference pipeline still require local PCA normals and covariance-derived planarity/curvature corrections? Please clarify what geometric information is actually internalized and what remains explicitly computed at test time.

Reviewer_GVf7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses the point cloud denoising issue by introducing a loss function aimed at separating errors caused by off-surface and on-surface misalignments. Additionally, a gating mechanism is responsible for suppressing the on-surface term in regions with sharp edges. The method is used in conjunction with SotA models on three separate datasets.

Critical strengths

Attempting to separate the loss contributions into off- and on-surface contributions is conceptually sound and interesting. Also, the gating mechanism is reasonable in principle.

Critical weaknesses

I have concerns regarding the experimental procedure and reporting, as well as some of the terms introduced in Eq. 8; further details are provided in the questions section.

Overall, I believe that the presentation of the paper should be significantly improved across multiple fronts. See a detailed breakdown below.

  1. In terms of overall structure, a background section to include in the main paper and describing the underlying architectures and optimisation procedures accompanying the newly introduced loss would have been particularly beneficial. Also, the main comparison against the SotA methods should have been better described and included in the main paper (more in the questions section).

  2. The writing quality is low, probably because of irresponsible or naïve use of AI assistance. While I appreciate the open declaration of its adoption, ultimately, it is the authors’ responsibility to guarantee sufficient quality standards. To be precise, terminology is often comprehensible yet imprecise or debatable (e.g., manifold recovery instead of the actual task of point cloud denoising, core velocity backbone instead of something like a lightweight model, etc.), multiple sentences are a vague combination of inflated terms (e.g., “inevitably leaks uncontrolled in-plane contraction force” -> why is it inevitable? Why is it uncontrolled?; “execution latency explosion”-> what does it even mean?). I would advise adopting terminology more consistent with the literature and keeping it consistent throughout the paper. Additionally, scrutinise every word generated by an LLM, and question whether it makes sense and is coherent with the literature, the general field, and the rest of the terminology established in the paper.

  3. Mathematical notation needs to be consistent and properly introduced; for instance, what is \bar{p}? Is it equivalent to \mu_i (line 214)?

  4. Figures are uninformative, which is sad to see in a paper having strong ties with the graphics community, where figures are always very curated. The purpose of a good scientific figure is to be informative, not merely aesthetic. I am certain that Figure 1 was AI-generated (the style is recognisable, circles are not perfect, text is rasterised, and parts of the illustration are generally meaningless), I am conscious that this does not require a mandatory disclosure, but my problem is with the image per se and not with how it was generated. 1A depicts gradients of loss terms that were not introduced in the paper; I am not sure what coupled dynamics and gradient interference are referred to as they currently point to randomly colored points. 1B Normal subspace is repeated twice; I don’t see any relation between the arbitrary manifold plot there, the ones in 1A, and the caption. What are the X shape and two bars supposed to be? In 1C bottom, the only thing blurred is the rendering, I am not even sure how a point cloud can be blurred. Here the authors could have really used examples from their results and SotA. Figure 2 is a dump of math rather than a simplification of the underlying idea and architecture. I am not sure it adds any value as it stands. Figure 3 shows results, but the first 5 columns are completely useless as they appear identical unless zooming beyond 500%, and even then only the small noise perturbation becomes visible. Put a small image of the object and show close-ups and cross sections instead. Additionally, a comparison against other SotA methods would have been useful. In Figure 4, I cannot appreciate any visible difference; maybe change the upper bound of the colourmap?

  5. Small additional note: the third contribution listed by the authors is not a contribution. Experiments are necessary to substantiate the claims of the paper, not a contribution. Results can be mentioned among the contributions but not as a contribution (e.g., we introduced A, which yielded X% better results). If Partition-of-Unity patch assembly was the main contribution meant to be highlighted at the third bullet point, it should be properly explained in the paper and not name-dropped throughout the paper without comprehensive explanation.

Questions for authors
  1. How are g_fit and g_uni defined (line 195) ? How can they be obtained?

  2. What is \bar{p}? Is it equivalent to \mu_i (line 214)?

  3. What is the distance head mentioned at line 265? Is it a neural network, a learnable parameter, or something else?

  4. V perpendicular is obtained by projecting v_\theta onto n*, therefore, it already shares the same direction as n*. After normalisation (first term of Eq.8), it also shares the same length. In light of this, I don’t understand why the first loss term should ever differ from 0. Similarly, since v parallel is always perpendicular to n* by construction, why should the first term of the second line in Eq. 8 ever differ from 0?

  5. How can you guarantee that n* is normal to the surface? I would argue that it is unlikely to have this property unless x_0 is the projection of x_1 onto the surface.

  6. Manifolds are repeatedly mentioned, but many CAD models aren’t manifold; does the method apply also to non-manifold objects?

  7. The Riemannian Manifold Hypothesis was introduced in latent representations, but point clouds sampled from or approximating 2D manifolds are known to come from there; it is not a hypothesis. Technically, the surface of these objects may not even be a manifold. I am rather confused as to why it is even mentioned, and it makes me question most of the remaining content of the paper.

  8. CAD models are also repeatedly mentioned, for instance at line 323 “high-density CAD models from ABCNet spanning 50k → 200k points…”, it is worth noting that the authors operate on point clouds sampled from CAD models, so the terminology is incorrect. The CAD models are not high-density; the sampled point clouds are.

  9. It would be useful to observe more qualitative and quantitative results with increased amounts of noise.

  10. The only proper comparison against SotA models is in Table 5 and Section D. It is unclear whether the reported metrics were obtained by reproducing the models or copied from previous papers. Were models matched in parameter count? Did they all run with fair and comparable parameters?

  11. According to Table 5, the proposed method is outperformed with low point counts; why?

  12. Table 1 compares the proposed method with StraightPCF, why this and why only this model? According to Table 5 StraightPCF was not the only suitable candidate for a comparison since also IterativePFN and 3DMambaIPF reported very strong performance metrics.

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

Related papers

Discussion (0)

Sign in to comment.