Pith. sign in

REVIEW 3 major objections 5 minor 3 references

Geometry Denoising with Preferred Normal Vectors

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A single variational problem can denoise a triangulated surface and segment it at the same time, by pushing each triangle's normal toward one of a prescribed set of preferred unit vectors and applying total-variation regularization to the l

desk verdict A clean, new coupling of denoising and label segmentation with an explicit ADMM/shape-Newton solver; the model is real, but the headline F1 gap is in-sample and the optimizer has no convergence guarantee. read the letter →

arxiv 2511.04848 v2 pith:D6GM2HCU submitted 2025-11-06 cs.CV math.OC

Manuel Wei{ss} , Lukas Baumgärtner , Roland Herzog , Stephan Schmidt This is my paper · ORCID
classification cs.CVmath.OC MSC 65D1849Q1049M1565K0590C30
keywords geometrydenoisingmeshsegmentationpreferrednormalvectorsassignmentfunctiontotalvariationADMMshapeNewtonsplitBregman
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes to solve geometry denoising and segmentation as one optimization: move mesh vertices so each triangle normal matches one of a given set of preferred normal vectors, while a total-variation term on the assignment encourages flat, contiguous regions. It claims the resulting nonconvex problem is tractable via ADMM, where all subproblems except the vertex update are closed-form or simple convex problems, and the vertex update is a second-order shape-Newton step. If the claim holds, applications with known preferred normals—urban LiDAR scans, crystal facets, carved surfaces—can be denoised substantially better than normal-blind total-variation denoising: in the city skyline experiment the proposed model reaches a distance of 0.138 to ground truth versus 0.755 for the baseline.

What carries the argument

The load-bearing object is the assignment function φ_T ∈ Δ_L on each triangle, which simultaneously encodes the segmentation and weights the normal-alignment penalty; its total variation (2.1) is the regularizer that creates flat regions of constant label. The ADMM scheme with auxiliary variables u, v, w decouples the non-smooth exact-penalty, TV, and simplex constraints, and the shape-Newton system (3.14) for vertex motion is the step that actually deforms the mesh toward the preferred normals.

What would settle it

Re-implement Algorithm 1 and run the city skyline experiment from multiple random initial assignments φ(0) (and random multiplier starts) with the paper's parameter values; if the spread of final distances to the ground truth is large, or if any run stops with a small variable change but an objective value far above the reported result, the claim that the scheme reliably solves (3.1) is not supported. On a small mesh with L = 4 labels, an independent check would be to enumerate all labelings and compare the ADMM output against the true global minimizer of (3.1).

Watch

Extended reading notes

Core claim

The central discovery is that denoising and segmentation can be coupled through a single assignment variable. The model minimizes F(Γ) + α Σ_T |T| Σ_ℓ φ_{T,ℓ} |n_T − g^{(ℓ)}|^2 + β TV(φ) over vertex positions and simplex-valued assignments φ_T; because the assignment penalty acts as an exact penalty, φ generically chooses one label per triangle, and the TV term groups triangles into patches of constant normal direction. The paper shows that an ADMM splitting makes each update tractable: the u, v, w subproblems are soft-thresholding or simplex projection, the φ-subproblem is a strongly convex sparse quadratic solved by conjugate gradients, and the x-subproblem is a globalized shape Newton ste

Load-bearing premise

The method's success depends on the assumption that the iterative solver lands on a good solution of a problem with many local solutions; the paper only checks that values stop changing between iterations, not that the solution is globally good.

Editorial extensions

If this is right

  • Urban mesh denoising from noisy LiDAR scans can exploit known axis-aligned building normals to recover flat facades instead of merely smoothing them.
  • In crystallography, where material symmetry restricts possible facet normals, the same model can segment grains or facets while cleaning the geometry.
  • The TV weight β directly controls the size of constant-label regions, so a user can trade fine detail against large flat patches by tuning one parameter.
  • With a large assignment weight α, the model can drive large geometric deformations (sphere to tetrahedron or dodecahedron), making it a shape stylization tool as well as a denoiser.
  • The baseline comparison indicates that when preferred normals are available but unused, denoising quality suffers dramatically; the gain is not incremental.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same ADMM template should transfer to other surface attributes, such as principal curvature directions or tangent vector fields, by replacing the normal n_T with the attribute of interest and keeping the assignment/TV structure unchanged.
  • A natural extension the paper leaves implicit is learning the label set from the noisy mesh itself—clustering normals first and then running the model—which would remove the need for an external prior in applications where preferred normals are not known exactly.
  • Because the exact-penalty character of the assignment term suggests φ converges to simplex vertices, a provable convergence rate would likely require explicit conditions linking α, β, and mesh quality; the paper does not provide such conditions.
  • The practical message from the parameter study is that α must dominate the noise level; too small an α makes the model degenerate toward normal-blind TV denoising, so a principled rule for choosing α from noise variance would be needed for automated use.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a variational model for simultaneous denoising and segmentation of triangulated surface meshes. The objective (1.1) combines a fidelity term F(Γ), an assignment term that aligns each triangle normal with one of a prescribed set of preferred unit normal vectors g^(ℓ) via a simplex-constrained label function φ, and a total-variation regularizer on φ. An ADMM scheme (Algorithm 1) is derived: the u, v, w subproblems have closed-form or projected updates, the φ-subproblem is a sparse quadratic solved by CG, and the vertex update is a shape-Newton step with truncated CG and line search. Experiments include a sphere, platonic solids, a synthetic city skyline, and the Stanford bunny. The headline result is that, on the city skyline, the proposed model recovers the ground truth with F1(Γ;x_true)=0.138 versus 0.755 for a TV-only baseline.

Significance. If the method is reliable, it provides a useful integration of shape optimization and discrete labeling with TV regularization, with potential applications in architecture and crystallography where preferred normals are known. The paper contains explicit derivations, a global-solution lemma for a nonconvex scalar subproblem, and reproducible parameter tables. However, the numerical central claim is weakened by in-sample hyperparameter tuning and the absence of convergence guarantees for the nonconvex ADMM scheme. The core model is defensible, but the evidence is not yet conclusive.

major comments (3)
  1. [Section 3.2, Eqs. (3.9)–(3.10)] The v-subproblem printed in (3.9) contains φ^{(k)}_{E+,ℓ} − φ^{(k)}_{E+,ℓ}, i.e., the same index twice, making the term identically zero. The update (3.10) repeats this error. The correct expression should use φ^{(k)}_{E+,ℓ} − φ^{(k)}_{E−,ℓ}, consistent with the constraint v_E = φ_{E+} − φ_{E−} and with the multiplier update in Algorithm 1, line 10. As written, Algorithm 1 would not couple adjacent triangles through the TV term, so the printed algorithm does not implement the model. This is a load-bearing typo and must be corrected.
  2. [Section 3.6, Algorithm 1] Problem (3.1) is nonconvex: φ multiplies u (thus the term α φ |n_T − g|²), and n_T depends nonlinearly on x. The paper defines convergence only as 'the absolute change in all eight variables is small between two successive iterations' — an empirical stopping criterion, not a convergence result. The x-subproblem is solved approximately with truncated CG and negative-curvature fallback, and the φ-subproblem uses CG with RTOL=10⁻², so standard ADMM convergence theorems (which typically require exact solves or a Kurdyka–Łojasiewicz property) are not verified. The reported improvements, including the city-skyline F1 advantage, depend on reaching a good stationary point. Please provide a convergence analysis under stated assumptions or, failing that, empirical evidence of convergence (e.g., objective and residual histories, multiple random initializations, sensitivity to tolerances) to support
  3. [Section 4.3] The hyperparameters α, β for the proposed model and γ for the baseline are chosen by manual grid search to minimize F1(Γ;x_true) on the same synthetic city instance, and the resulting F1 is then reported as the outcome. This is an in-sample evaluation; the headline comparison (0.138 vs 0.755) may reflect parameter tuning rather than intrinsic model superiority. Only a single noisy instance is tested. Please evaluate on held-out noise realizations or a small test set, and report how the F1 varies over a range of α, β, and γ (e.g., a table or a contour plot). At minimum, state explicitly that the result is in-sample and justify that the selected parameters are representative.
minor comments (5)
  1. [Section 4.3, Figure 4.5d] The text states that the baseline model uses γ=0.015, but the caption of Figure 4.5d says 'β=0.015'. Please unify the notation with the baseline model (1.2).
  2. [Table 4.1 caption] The caption 'Figures 4.2a to 4.2c 4.2d to 4.2f 4.2g 4.2h 4.2i' is malformed; please reformat.
  3. [References] The URL for the Humboldt University page contains a typo: 'httpss://www.mathematik.hu-berlin.de/...'. Also, check whether 'mshr' and 'meshio' should be cited with appropriate references.
  4. [Section 3.4] Stating that the linear system is solved using 'the PETSc implementation of the conjugate gradient method' is slightly imprecise; PETSc is a library that provides CG. Consider wording.
  5. [Section 4.2] The tetrahedron and dodecahedron have 4 and 12 face normals respectively; the text says 'L=4' and 'L=12' but does not specify how these normals are chosen for a sphere-initialized mesh. A brief clarification would help.

Circularity Check

1 steps flagged · score 4.0 of 10

Main modeling claim is not circular, but the city-skyline comparison is in-sample: hyperparameters are tuned on the same F1 metric that is then reported as the method's outcome.

  1. fitted input called prediction [Section 4.3, 'City Skyline Denoising' (parameter selection and reported F1 values)]
    "The parameters α and β for (3.1) and parameter γ for the baseline model (1.2) are chosen by a manual grid search for each model such that the resulting mesh approximately minimizes F1(Γ;x_true) as a measure of distance to the ground truth mesh. ... The proposed model (3.1) can recover the ground truth mesh with F1(Γ;x_true) = 0.138, whereas the baseline TV denoising model without any additional normal direction information (1.2) results in a distance of F1(Γ;x_true) = 0.755"

    The reported quantitative superiority is not an independent model-derived prediction: the hyperparameters α, β, γ of both models were selected by grid search to minimize the exact same F1(Γ;x_true) objective that is then reported as the outcome. The headline numbers are therefore in-sample best-case values of the selection procedure rather than a test of how the method behaves for fixed, a priori chosen parameters. This makes the numerical comparison partly circular, although the variational model itself is not defined in terms of F1 and retains independent content.

full rationale

No other circular step is present. Model (1.1)/(3.1) is proposed as a new combination of a normal-TV denoising fidelity and a simplex-constrained assignment TV term, and the derivation of the ADMM updates (3.8), (3.10), (3.12), (3.13) and the shape-Newton step (3.14) follows from the stated augmented Lagrangian (3.3), not from the conclusions. Citations to the authors' own prior work supply the baseline (1.2), the segmentation prototype (1.3), and the shape-Newton fallback details; these are legitimate method references and are not used to derive the claimed result by fiat. The convergence criterion in Section 3.6 is only an empirical stopping rule, and the v-subproblem (3.9)/(3.10) contains an apparent φ_{E+}−φ_{E+} typo, but these are correctness/robustness limitations, not circularity. The central modeling contribution is self-contained; the main circularity concern is confined to the city-skyline benchmark being tuned on its own evaluation metric.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method's central claim rests on a user-supplied prior (finite label set), a smooth fidelity term, and the heuristic convergence of a nonconvex ADMM scheme. The free parameters are regularization and algorithm weights, all chosen by hand or grid search; none are estimated within the method, so the method itself is not a fitted model, but the reported quantitative comparison is affected by test-set tuning.

free parameters (6)
  • assignment weight α = 0.1–100 (per experiment)
    Weights the normal-alignment term in (1.1); chosen by hand in Tables 4.1–4.4 and by grid search against ground truth in Section 4.3.
  • TV weight β = 1e-8 to 0.1
    Weights TV(φ); controls size of constant-label regions; chosen by hand/grid search; in Section 4.3 tuned using x_true.
  • mesh quality weight κ = 1e-11 to 1e-5
    Weight of area penalty F2 in (4.1); prevents triangles from becoming too small; chosen by hand per experiment.
  • ADMM augmentation parameters ρ1, ρ2, ρ3 = 0.2–10000 per experiment
    Penalty parameters in augmented Lagrangian (3.3); chosen by hand per experiment.
  • inner-product parameter c = 0.1 or 0.3
    Preconditioner parameter in (3.15); chosen by hand per experiment.
  • number of labels L and label directions g^(ℓ) = L=4,6,12,20,30 in experiments
    User-supplied prior; in each experiment chosen to reflect expected geometry (axis directions, platonic normals, Fibonacci lattices). Not estimated from data.
assumptions (5)
  • domain assumption The mesh has fixed connectivity, is manifold, and remains non-degenerate during optimization.
    Used throughout Section 2 and in Section 3.5 (“Provided that the mesh remains non-degenerate (i.e., a manifold), which we monitor”); the shape Newton step presupposes smooth dependence on vertex positions.
  • domain assumption The fidelity term F is smooth in vertex positions.
    Eqs. (4.1)–(4.2): F1 is quadratic in vertex positions; F2 = sum 1/|T| is smooth away from degenerate triangles.
  • domain assumption The user supplies a finite set of unit preferred normals containing the directions present on the clean surface.
    This is the prior-knowledge premise of the method, stated in Section 1 (“we can expect to possess prior knowledge about preferred normal vectors”).
  • ad hoc to paper The alignment term acts as an exact penalty, driving φ toward simplex vertices and normals toward labels for large α.
    Section 1 asserts “This is due to it acting as an exact penalty function”; no proof is given for the TV-coupled problem, and in practice α remains finite (α=20 for platonic solids).
  • ad hoc to paper The ADMM iteration converges to a useful stationary point of the nonconvex problem.
    Section 3.6 defines stopping by small absolute changes; no convergence or rate analysis is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Geometry Denoising with Preferred Normal Vectors." pith.science (2026). https://pith.science/paper/D6GM2HCU

@misc{pith2026251104848,
  author       = {Pith},
  title        = {Pith review of: Geometry Denoising with Preferred Normal Vectors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D6GM2HCU}},
  note         = {Machine review of arXiv:2511.04848}
}
read the original abstract

We introduce a new paradigm for geometry denoising using prior knowledge about the surface normal vector. This prior knowledge comes in the form of a set of preferred normal vectors, which we refer to as label vectors. A segmentation problem is naturally embedded in the denoising process. The segmentation is based on the similarity of the normal vector to the elements of the set of label vectors. Regularization is achieved by a total variation term. We formulate a split Bregman (ADMM) approach to solve the resulting optimization problem. The vertex update step is based on second-order shape calculus. We present various examples including the denoising of an eroded medieval gravestone inscription.

Figures

Figures reproduced from arXiv: 2511.04848 by the authors.

Figure 4
Figure 4. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 4.1
Figure 4.1. Noisy input data for the sphere geometry experiment (Section 4.1) [PITH_FULL_IMAGE:figures/full_fig_p010_4_1.png] view at source ↗
Figure 4
Figure 4. i [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figures from the paper (8 more)
Figure 4.2
Figure 4.2. Figure 4.2: Solutions for problem (3.1) for the sphere example (Section 4.1, L = 20 labels) with noise as shown in [PITH_FULL_IMAGE:figures/full_fig_p011_4_2.png]
Figure 4
Figure 4. Figure 4: a [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 4.3
Figure 4.3. Figure 4.3: Iteration history of Algorithm 1 for the tetrahedron platonic solid problem (Section 4.2, L = 4 labels) with a sphere as initial guess and an assignment weight α = 20. Cells are colored according to the assigned label. The results of this experiment are shown in [PI…
Figure 4.4
Figure 4.4. Figure 4.4: Iteration history of Algorithm 1 for the dodecahedron platonic solid problem (Section 4.2, L = 12 labels) with a sphere as initial guess and an assignment weight α = 20. Cells are colored according to the assigned label. 4.4. Stanford Bunny. In this final example, we…
Figure 4
Figure 4. Figure 4: c [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 4.5
Figure 4.5. Figure 4.5: Denoising of the city skyline geometry (Section 4.3) with the proposed model (3.1) (bottom left) and the baseline total variation-based model (1.2) (bottom right), which is not informed by preferred normal directions. labels, respectively, distributed uniformly aroun…
Figure 4
Figure 4. Figure 4 [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 4.6
Figure 4.6. Figure 4.6: Solutions for problem (3.1) for the Stanford bunny and different numbers L of preferred normal vectors. Cells are colored according to the assigned label. 5. Conclusion In this paper, we presented a novel model (3.1) for denoising triangular surface meshes based on a…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 3 linked inside Pith

  1. [1]

    The FEniCS project version 1.5

    Alnæs, M.; J. Blechta; J. Hake; A. Johansson; B. Kehlet; A. Logg; C. Richardson; J. Ring; M. E. Rognes; G. N. Wells (2015). “The FEniCS project version 1.5”. Archive of Numerical Software3.100, pp. 9–23.doi:10.11588/ans.2015.100. 20553. Balay, S.; S. Abhyankar; M. Adams; J. Brown; P. Brune; K. Buschelman; E. Con- stantinescu; A. Dener; J. Faibussowitsch; ...

    arXiv 2015

  2. [13]

    Total variation regularization for functions with values in a manifold

    Lellmann, J.; E. Strekalovskiy; S. Koetter; D. Cremers (2013). “Total variation regularization for functions with values in a manifold”.2013 IEEE International Conference on Computer Vision, pp. 2944–2951.doi:10.1109/ICCV.2013.366. Lou, Y.; M. Yan (2017). “Fast L1-L2 minimization via a proximal operator”.Journal of Scientific Computing74.2, pp. 767–785.do...

    arXiv 2013

  3. [2002]

    A signal processing approach to fair surface design

    IEEE, pp. 125–132.doi:10.1109/VISUAL.2002.1183766. Taubin, G. (1995). “A signal processing approach to fair surface design”.Proceedings of the 22Nd Annual Conference on Computer Graphics and Interactive Tech- niques. SIGGRAPH ’95. ACM, pp. 351–358.doi:10.1145/218380.218473. Wang, W.; M. Á. Carreira-Perpiñán (2013).Projection onto the probability simplex: ...

    arXiv 2002

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.