{"id":"caf4d8fc-5ec7-46d3-8bf1-2fb7bf0a0d7a","arxiv_id":"2511.04848","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Mesh denoising and segmentation are combined in one optimization that aligns triangle normals with prescribed preferred directions under total-variation regularization.","lead":"This paper proposes a variational model that denoises 3D triangle meshes by pushing each triangle's normal toward one of a user-supplied set of preferred directions, while simultaneously labeling each triangle by the direction it follows. It matters for reconstructing noisy scans of buildings, crystals, and other surfaces whose orientations are known ahead of time.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 has no convergence guarantee for the nonconvex problem (3.1); the reported F1 advantage may depend on initialization and in-sample tuning.","rationale":"The reader's weakest_assumption was exactly the lack of convergence analysis for the nonconvex ADMM/shape-Newton iteration. My reading of the paper finds no additional flaw that supersedes this. The typos in (3.9)/(3.10) are real and would break a verbatim implementation, but they are easily corrected and do not invalidate the underlying method. The in-sample tuning of hyperparameters on ground truth is a methodological weakness but does not by itself undermine the model's plausibility. The absence of code makes the convergence concern harder to resolve, but the proposed concrete test can settle it. Therefore the paper remains CONDITIONAL: the central modeling idea is coherent and the experiments qualitatively support it, but the numerical claims require reproducible, initialization-independent evidence.","tokens_in":14164,"tokens_out":6127,"duration_ms":57255,"concrete_test":"Re-implement Algorithm 1 with the corrected v-update (using φ_{E+}−φ_{E-}). Run the city-skyline experiment (Section 4.3) from 10 random initializations of φ⁰, μ⁰, ν⁰, λ⁰ (or random perturbations of the noisy input). Record the final F1(Γ;x_true) and the final objective value. If the variance across runs is small and the KKT residual (or proximal gradient norm) at the final iterate is below a small tolerance, the convergence concern is mitigated. If the solutions vary significantly or the stationarity residual is large, the reported F1 gain is not reproducible and the central numerical claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that problem (3.1) can be solved by Algorithm 1 and that this yields superior denoising/segmentation, e.g., the city-skyline F1=0.138 vs 0.755. The weakest load-bearing assumption is that the nonconvex ADMM/shape-Newton iteration reaches a good stationary point. Problem (3.1) is nonconvex because φ multiplies u (and hence |n_T−g|²) and n_T depends nonlinearly on x. Section 3.6 defines convergence only as 'the absolute change in all eight variables is small between two successive iterations' — an empirical criterion, not a proof of convergence to a stationary point. For nonconvex ADMM, convergence to a stationary point typically requires the augmented Lagrangian to satisfy the Kurdyka-Łojasiewicz property and exact or sufficiently accurate subproblem solves. Here the x-subproblem is solved only approximately via truncated CG with negative-curvature fallback, so those conditions are not verified. Consequently, the reported solutions may be initialization-dependent poor local minima, and an independent implementation may not reproduce them. Additionally, the v-update (3.10) contains a typo: it writes φ_{E+}−φ_{E+} instead of φ_{E+}−φ_{E-}, making the printed algorithm incorrect if implemented verbatim. The city-skyline hyperparameters are manually tuned on the ground-truth mesh, so the headline comparison is in-sample. These factors together make the numerical central claim fragile.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":14572,"tokens_out":4010,"duration_ms":41064,"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":[{"comment":"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.","section":"Section 3.2, Eqs. (3.9)–(3.10)"},{"comment":"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","section":"Section 3.6, Algorithm 1"},{"comment":"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.","section":"Section 4.3"}],"minor_comments":[{"comment":"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).","section":"Section 4.3, Figure 4.5d"},{"comment":"The caption 'Figures 4.2a to 4.2c 4.2d to 4.2f 4.2g 4.2h 4.2i' is malformed; please reformat.","section":"Table 4.1 caption"},{"comment":"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.","section":"References"},{"comment":"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.","section":"Section 3.4"},{"comment":"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.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a reasonable combination of known ingredients, but the numerical evaluation and convergence guarantees are not yet at the level expected for acceptance. The typos in Eqs. (3.9)–(3.10) should be caught in any careful revision. If the authors can demonstrate algorithmic convergence empirically and provide out-of-sample evaluation, the paper could become a solid contribution. The novelty is moderate, but the application domain and practical relevance are clear."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe new thing here is model (1.1): one variational problem that couples mesh denoising with a normal-label assignment, regularized by TV on the assignment. The authors are honest that it combines their own earlier denoising model (1.2) and segmentation model (1.3), and the combination is genuinely new and well motivated. The ADMM derivation is explicit, including a lemma for the nonconvex u-update that works for negative coefficients, and the shape-Newton subproblem is a reasonable continuation of their SISC work. The self-citations are appropriate here; this is a direct line of work. The experiments show the intended qualitative behavior: larger α aligns normals with labels, larger β grows constant-label regions, and the city-skyline comparison shows a large quantitative improvement over the TV-only baseline.\n\nThe quantitative claim, though, is weaker than it looks. In Section 4.3 the hyperparameters are chosen by a manual grid search that approximately minimizes F1(Γ; x_true), and the same metric is then reported as the outcome. That makes the 0.138 vs 0.755 comparison in-sample, not a prediction. There is also no convergence proof for the nonconvex ADMM; Section 3.6 merely checks that the variables stop changing. I do not think that is fatal for a methods paper — the iterations visibly work on four examples — but it means an independent implementation could land at a different stationary point, and the typo in (3.9)–(3.10), with φ_{E+} repeated instead of φ_{E+} − φ_{E−}, would actually break the v-update if implemented verbatim.\n\nThe evaluation is also thin: one noisy instance per experiment, no external baseline, no code or data. For a paper whose main evidence is numerical, that is a real soft spot.\n\nWho gets value: someone working in mesh denoising, shape optimization, or CAD/LiDAR reconstruction will find this a useful template. It is not a breakthrough, but it is a clean and citable methods contribution.\n\nI would send it to peer review rather than desk reject. The referee should ask for code, a proper validation protocol, and a correction of the typo; with those in place the paper is solid. I would cite it if I worked on this class of problems.","headline":"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.","tokens_in":15066,"tokens_out":4860,"would_cite":true,"duration_ms":38906,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D18","49Q10","49M15","65K05","90C30"],"pacs":[],"model":"deepseek-v4-flash","headline":"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","keywords":["geometry denoising","mesh segmentation","preferred normal vectors","assignment function","total variation","ADMM","shape Newton","split Bregman"],"falsifier":"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).","tokens_in":14025,"feed_emoji":"🏙️","tokens_out":6908,"duration_ms":65155,"temperature":0.7,"pith_summary":"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.","feed_headline":"Preferred normals cut mesh denoising error fivefold","feed_subtitle":"On a noisy skyline scan, it recovers the true geometry with about five times less error than normal-blind denoising.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Segment and denoise meshes with label vectors, 5x less error","Preferred normals slash mesh denoising error fivefold via ADMM","Geometry denoising gets 5x better by embedding segmentation","Label-vector denoising couples segmentation for 5x accuracy","Mesh denoising with normal priors: 5x lower error, one pass"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Segment and denoise meshes with label vectors, 5x less error","Preferred normals slash mesh denoising error fivefold via ADMM","Geometry denoising gets 5x better by embedding segmentation","Label-vector denoising couples segmentation for 5x accuracy","Mesh denoising with normal priors: 5x lower error, one pass"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":997,"prompt_tokens":651,"completion_tokens":346,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":395,"completion_tokens_details":{"reasoning_tokens":249}},"tokens_in":395,"tokens_out":346,"duration_ms":4255,"temperature":1.0,"reasoning_tokens":249,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T23:34:09.087480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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).","supporting_citations":[],"review_version":1}