{"id":"978ac6c0-0cea-4702-b93c-9fe064d3d690","arxiv_id":"2411.14158","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GD-GCN denoises 3-D point clouds with micro-step graph convolution, a learned Riemannian graph, and Bernstein spectral filters, reporting lower CD, EMD, and HD than existing methods on several benchmarks.","lead":"A new point cloud denoising network, GD-GCN, passes messages between points in many small ODE-style steps and builds its graph with a learned distance metric. It reports lower Chamfer and Earth mover distances than ten baselines on synthetic and LiDAR benchmarks, but releases no code and its geometric derivations contain gaps.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (8)'s attention output cannot be the per-point Jacobian required by Eq. (7): the dimensions mismatch, so the learned 'Riemannian metric' is unsupported and the geometric-graph contribution loses its stated basis.","rationale":"The reader's weakest-assumption analysis correctly identifies Eq. (8) as the point where the geometric contribution is least secure. The dimension inconsistency is not a matter of stylistic notation: a Jacobian is intrinsically per-point and has mixed input/output dimensions, while the attention output is either an N x N score matrix or, after linear projection, a single N x d matrix. No amount of reinterpreting 'approximation' bridges this gap without an explicit construction and error analysis. Because the geometric graph is one of the paper's three advertised mechanisms, an unsupported metric claim weakens the explanatory version of the central claim. I give credit where it is due: the empirical comparisons and ablations are reported consistently, and the micro-step and Bernstein-filter contributions may well be useful independent of the geometric framing. But the lack of released code, error bars, or a derivation for Eq. (8) means the empirical results cannot be independently checked and the theory cannot be repaired by inspection. These are addressable issues, not fatal flaws, so the reader's CONDITIONAL verdict remains appropriate; my read does not move it, hence verdict_should_be is UNCHANGED.","tokens_in":21068,"tokens_out":4924,"duration_ms":49475,"concrete_test":"Implement Eqs. (6)-(8) exactly with the actual layer dimensions d and d_l and execute one forward pass on a 1024-point patch, printing the tensor shapes. If J_theta is not a per-point matrix of shape d x d_l (or d_l x d) after the described concatenation and linear transform, Eq. (8) cannot approximate the Jacobian and the Riemannian-metric claim fails as stated. Separately, rerun the Geo-Graph ablation while retaining the identical attention block but replacing the quadratic-form distance with the Euclidean distance used in 'GD-GCN w/o Geo-Graph'; if the CD gap shrinks to near zero relative to the reported Fig. 8 values (1.550 vs. 1.637 x 10^-6, about 5.6%), then the observed benefit is attributable to added model capacity rather than to Riemannian geometry.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim names three mechanisms for outperforming prior art: micro-step temporal graph convolution, Riemannian-metric graph construction, and Bernstein spectral filtering. The load-bearing weak point is the second mechanism. In Section III-D2, Eq. (8) defines J_theta as the output of a multi-head attention expression, J_theta = LeakyReLU( X_tilde W_Q (X_tilde W_K)^T / sqrt(d) ). For X_tilde in R^{N x K(d+d_l)}, the product (X_tilde W_Q)(X_tilde W_K)^T has shape N x N, not the per-point Jacobian shape required by Eq. (7). Even after head concatenation and a linear projection to R^{N x d}, the result is a single matrix shared across points, whereas the Jacobian of z(p) is a per-point linear map with shape d x d_l for z: R^{d_l} -> R^d. Eq. (7) itself is also dimensionally inconsistent with the stated map, since the product (nabla z(p))^T (nabla z(p)) yields a d_l x d_l matrix when z maps R^{d_l} to R^d, not the required d x d metric. Consequently, the 'Riemannian distance' computed in Eq. (6) is at best a learned quadratic form with no demonstrated relation to the manifold geometry. The ablation comparing geometric and Euclidean graphs therefore cannot support the paper's geometric claim: it only shows that replacing Euclidean edge weights with this attention-based construction changes the output. The paper provides no derivation, no error bound, and no code to resolve the discrepancy, making this the most load-bearing concern for the stated contributions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GD-GCN, a graph-convolutional network for 3-D point cloud denoising. The method has three advertised components: (i) MST-GConv, a micro-step temporal graph convolution interpreted through the lens of neural ODEs/PDEs; (ii) a geometric graph construction that claims to approximate a Riemannian metric from an attention-based Jacobian estimate; and (iii) a Bernstein-polynomial graph spectral filter with a claimed BIBO-stability guarantee and symmetric channel mixing for spectral adaptation. The authors report supervised and unsupervised experiments on ShapeNet, Stanford, and Paris-rue-Madame data, with CD, EMD, HD, and RMSD metrics, and they provide ablation studies comparing the proposed graph construction and spectral filtering against simplified variants. The central claim, stated in the abstract and Section I, is that GD-GCN outperforms state-of-the-art denoising methods through these three mechanisms.","tokens_in":1723,"tokens_out":4483,"duration_ms":112767,"significance":"If the technical gaps were repaired, GD-GCN would be a useful contribution: the neural-ODE perspective on dynamic graph construction is timely, the experimental campaign is broad (synthetic and LiDAR data, supervised and unsupervised settings), and the reported CD/EMD gains in Tables II-V are substantial. The paper also includes an ablation study that isolates the proposed components, which is a strength. However, the Riemannian-metric derivation and the BIBO-stability proof contain load-bearing errors, and the empirical claims are presented without error bars or code. The significance is therefore conditional on correcting the theoretical claims and strengthening the experimental evidence.","major_comments":[{"comment":"The learned object J_theta is not a Jacobian in any stated sense. With X_tilde in R^{N x K(d+d_l)}, the product (X_tilde W_Q)(X_tilde W_K)^T is N x N before the concatenation and linear projection to R^{N x d}; there is no per-point d x d_l matrix. Moreover, Eq. (7) is dimensionally inconsistent with the stated map z : R^{d_l} -> R^d: (nabla z(p))^T (nabla z(p)) is d_l x d_l, not d x d. Consequently, the 'Riemannian distance' in Eq. (6) is at best a learned quadratic form, and the ablation in Section IV-D2 cannot establish the geometric claim. The authors should either provide a correct derivation of a per-point metric from a bona fide Jacobian, or explicitly downgrade the contribution to a learned Mahalanobis-style distance.","section":"Section III-D2, Eq. (8)"},{"comment":"The coefficient formula in Proposition 1 contradicts the normalization used in the proof. Taking g equiv 1 and K = 2 gives theta_k = 2^{-2} * 3! / 1! = 3/2 for k = 0, 1, 2; then B_2(lambda) = (3/2)[(1-lambda)^2 + 2 lambda (1-lambda) + lambda^2] = 3/2, so B_2(0.5) = 1.5 > 1, violating the claimed bound 0 < B_K(lambda) <= 1. The proof's normalization factor (K - floor(K/2))! / (K+1)! * 2^K is inverted relative to the stated theta_k; the correct coefficient would multiply by 2^K (K - floor(K/2))! / (K+1)!, not divide by it. As stated, the BIBO-stability guarantee is false and must be corrected.","section":"Proposition 1 and Appendix A"},{"comment":"The eigenvalue identity for channel mixing is dimensionally inconsistent. With mu, phi in R^d (eigenvalues of the d x d mixing matrices) and phi in R^N (eigenvalues of B_K), the eigenvalues of a Kronecker-sum operator acting on the vectorized N x d state lie in R^{dN} and are of the form mu_i phi_j - phi_k (or, under additional commuting assumptions, mu_i phi_j - phi_i), i.e., mu ⊗ phi - 1 ⊗ phi in the commuting case. The expression mu ⊗ phi - phi ⊗ 1 combines vectors of lengths d^2 and dN and cannot be the spectrum of the stated operator. Please correct the identity and the subsequent 'outer product' description, and state any simultaneous-diagonalizability assumptions needed for the Kronecker-sum eigenvalue factorization.","section":"Section III-E2, Eq. (11)"},{"comment":"The central empirical claim rests on Tables II-V, but no error bars, confidence intervals, or significance tests are provided, and the description of baseline training ('train the remaining models to achieve best performance', Section IV-A3) makes it difficult to assess whether the reported gains are robust or favorable to the proposed method. Reporting variance over at least a few random seeds or test subsets is necessary to support the 'outperforms state-of-the-art' conclusion, especially where the reported differences are small (e.g., the dense 30K cases at several noise levels in Table II).","section":"Section IV, Tables II-V"}],"minor_comments":[{"comment":"The abstract contains an incomplete sentence: 'fitting the point cloud with noise to the underlying surface by and the learning process for MST-GConv acts like a changing system' should be rephrased to remove the dangling 'by'.","section":"Abstract"},{"comment":"The word 'categoried' should be 'categorized'.","section":"Section II-B"},{"comment":"The sentence 'Subsequently, we the framework also updates A_t^G in real-time' is grammatically incomplete; it should be 'Subsequently, the framework also updates A_t^G in real-time'.","section":"Section III-C2"},{"comment":"The notation in the loss definitions is inconsistent: Eq. (13) mixes u_j and the denoised point variable in the same expression, and Eq. (14) writes a sum over points in the set of denoised points using an unclear subscript. Please align the indices carefully.","section":"Section III-F1, Eqs. (13)-(14)"},{"comment":"References [28] and [68] are the same paper (Wang et al., 'Dynamic graph CNN for learning on point clouds') listed twice with different numbers; please merge the duplicate citation.","section":"References"},{"comment":"The proposition states 'for any k in N' but k ranges over 0, ..., K; the quantifier should be 'for k = 0, 1, ..., K'.","section":"Proposition 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical study is extensive, but the two central theoretical contributions - the Riemannian-metric graph construction and the BIBO-stability guarantee - contain errors that are load-bearing for the stated claims. Both are correctable in principle: the metric can be reframed as a learned quadratic form, and the Bernstein coefficient normalization can be fixed. Given the scope of the changes needed, a major revision is appropriate rather than an outright rejection. The lack of error bars and code should also be addressed in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful core here is the architecture combination: continuous micro-step graph convolution (graph neural ODE) adapted to point cloud denoising, a learned attention-based graph weighting, and Bernstein polynomial spectral filtering. That combination is genuinely new for this task, and the experimental work is thorough. The paper reports consistent wins over ten baselines on ShapeNet, Stanford, and real LiDAR data, in both supervised and unsupervised modes, with ablations that support each component. The empirical tables look believable, and the margins, while incremental, are real. I give the authors credit for running a wide comparison and including real-world data.\n\nThe soft spots are concentrated in the theory, and they are not cosmetic. The Riemannian metric claim is the load-bearing one. Equation (8) defines J_theta as an attention output with shape roughly N x N (or N x d after projection), but Eq. (7) needs a per-point Jacobian of shape d x d_l for a map z: R^{d_l} -> R^d. The dimensions do not match, and Eq. (7) itself is inconsistent: (∇z)^T(∇z) would be d_l x d_l, not d x d. So the \"learned Riemannian metric\" is at best a learned quadratic form with no demonstrated geometric content. The ablation that replaces this with a Euclidean graph only shows that the attention weighting helps, not that any manifold geometry is being captured. This needs to be fixed or the claim needs to be walked back.\n\nThe Bernstein stability proof is also shaky. The normalization factor in Appendix A looks off by a factor of (K+1), and the proof shows existence of positive coefficients rather than bounding the learned coefficients actually used. That weakens the BIBO stability guarantee, though it is a secondary point.\n\nOn the empirical side: no error bars, no code, and vague baseline training descriptions (\"train the remaining models to achieve best performance\"). That makes the numbers hard to verify and is a legitimate knock on reproducibility.\n\nWho is this for? Anyone working on graph-based point cloud denoising or neural ODEs for geometry. The empirical contribution is useful even if the theoretical narrative needs repair. I would send this to peer review: the method and results deserve scrutiny, and the flaws are addressable. I would not cite the Riemannian metric story in its current form, but I would cite the empirical architecture after the authors clarify or retract the geometric claim.","headline":"Solid empirical architecture paper whose Riemannian-metric story doesn't survive contact with the equations, but whose denoising results still deserve review.","tokens_in":21976,"tokens_out":2109,"would_cite":true,"duration_ms":22098,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that denoising 3-D point clouds should be treated as a continuous dynamical system, and that micro-step temporal graph convolution, a learned Riemannian metric, and stable Bernstein spectral filters let one network beat…","keywords":["point cloud denoising","graph convolutional network","micro-step temporal graph convolution","neural ODE","Riemannian metric","Bernstein polynomial","graph spectral filtering","dynamic graph"],"falsifier":"Train GD-GCN, then on a held-out noisy point cloud compute the finite-difference Jacobian of the map from input 3-D coordinates to the $l$-th layer feature $z(p)$ for a set of points, and compare its action on neighbor displacement vectors with the attention-based $J_\\theta$ used in Eq. (8); a systematic mismatch (wrong dimension, sign flip, or large directional error) would falsify the claim that the graph construction is geometric.","tokens_in":20841,"feed_emoji":"🧊","tokens_out":7428,"duration_ms":61405,"temperature":0.7,"pith_summary":"The paper tries to establish that point-cloud denoising is best modeled as a continuous-time dynamic process, not as a stack of discrete graph convolutions. It proposes GD-GCN, which splits each message-passing step into many micro steps governed by a neural ODE, builds its graph with a learned approximate Riemannian metric instead of plain Euclidean distance, and filters in the graph spectral domain with Bernstein polynomials whose coefficients are constrained so the system stays bounded. The payoff, if correct, is that one network beats ten state-of-the-art denoisers on most standard noise levels in both supervised and unsupervised modes, with the largest reported gain at 1% noise on sparse clouds, and generalizes to real LiDAR data. That matters because denoising is a prerequisite for surface reconstruction, rendering, and downstream perception, and the method claims all this without per-noise-level retuning.","feed_headline":"Micro-step graph convolution beats prior point-cloud denoisers","feed_subtitle":"Splitting message passing into fine time steps plus a learned Riemannian metric lowers error on most noise levels.","key_machinery":"The load-bearing object is MST-GConv, a residual graph convolution written as a dynamical system: $Z^{l} = Z^{l-1} + \\delta((A_G^l - I)Z^{l-1}\\Theta^l)\\tau^l$ with micro-step size $\\tau^l \\ll 1$, solved numerically as a neural ODE or neural PDE (Eqs. 3–5). Around it are two supporting mechanisms: (1) geometric graph construction via the approximate Riemannian metric $G = I + \\alpha^2(\\nabla z(p))^\\top(\\nabla z(p))$ (Eq. 7), where the Jacobian is replaced by a multi-head attention output $J_\\theta$ (Eq. 8) and edge weights are $\\tilde{a}_{ij} = \\exp(-|\\ell_{ij}|^2/(2\\delta^2))$; and (2) a graph spectral filter built on the Bernstein polynomial basis $B_K(\\lambda) = \\sum_{k=0}^K \\theta_k \\binom{K}{k}(1-\\lambda)^{K-k}\\lambda^k$ with coefficient constraints giving $0 < B_K(\\lambda) \\le 1$ (Proposition 1, Eq. 9), combined with symmetric channel mixing matrices $W_1^t, W_2^t$ that control per-channel frequency scaling and shifting (Eqs. 10–11). The machinery's role is to let the network capture topology changes during denoising, separate geometric regions, and keep the evolution numerically stable while avoiding low-frequency dominance.","core_discovery":"The paper's central claim is that point cloud denoising should be treated as a continuous-time dynamical system rather than as a stack of discrete graph convolution layers. It introduces micro-step temporal graph convolution (MST-GConv), which replaces the one-hop message passing of a residual GCN with many small updates governed by a neural ODE, so that noisy points evolve smoothly toward the underlying surface. To build the graph for these updates, the method learns an approximate Riemannian metric $G = I + \\alpha^2(\\nabla z(p))^\\top(\\nabla z(p))$ and measures edge weights by the induced distance, separating points whose Euclidean proximity is deceptive (e.g., across sharp edges). To keep the evolution stable, it applies a Bernstein-polynomial spectral filter $B_K(\\lambda)$ with coefficients constrained so that $0 < B_K(\\lambda) \\le 1$, giving bounded-input bounded-output stability, plus symmetric channel mixing matrices that scale and shift frequency components to avoid low-frequency dominance. On supervised and unsupervised denoising benchmarks, the paper reports lower Chamfer distance and Earth Mover's distance than ten baselines across most noise levels, for example Chamfer distance $4.39 \\times 10^{-5}$ versus the previous best $6.02 \\times 10^{-5}$ on sparse 10K point clouds with 1% Gaussian noise.","pith_inferences":["The same micro-step temporal graph convolution could be transplanted to other graph-based geometry tasks (segmentation, upsampling, normal estimation), since the continuous-time formulation is task-agnostic; a testable extension would be replacing the point-coordinate signal with other features.","If the attention-based Jacobian truly approximates the Riemannian metric, the geometric graph construction should improve any distance-based point-cloud method, not just this architecture; an ablation that swaps the learned metric for a frozen Euclidean metric on the same backbone would isolate that contribution further.","The stability bound suggests the network can be integrated for longer times or with more micro-steps than tested, which the paper does not push to its limit; increasing $T$ beyond 1 at high noise levels would reveal whether the bound translates to better asymptotic fitting.","The paper's own results show the unsupervised version loses to the supervised one at 3% noise on sparse clouds; this hints that the learned metric and spectral filter may need noise-level-aware adaptation, a direction the authors do not explore."],"forward_implications":["On the 10K sparse test set at 1% Gaussian noise, GD-GCN reports Chamfer distance $4.39 \\times 10^{-5}$ versus $6.02 \\times 10^{-5}$ for the best prior method, and it leads on Earth Mover's distance as well.","At termination times $0.25T$, $0.5T$, $0.75T$, and $T$, the extracted intermediate states show the denoised cloud progressively approaching the ground-truth surface, confirming the continuous-evolution interpretation.","The Bernstein-polynomial filter bounds the filtered eigenvalues to $(0,1]$, so the dynamic system is BIBO stable and cannot diverge or amplify noise during long integration.","On real Paris-rue-Madame LiDAR data, the method achieves lower CD, HD, EMD, and RMSD than the compared unsupervised models, indicating that the learned dynamics carry over to realistic acquisition noise."],"supporting_citations":[{"why":"Supplies the neural-ODE formulation used to solve the micro-step temporal graph convolution as a continuous-time system.","marker":"[41]"},{"why":"Provides the Bernstein polynomial basis used to build the stable graph spectral filter.","marker":"[52]"},{"why":"The graph-convolutional denoising baseline whose residual layer scheme MST-GConv generalizes and which GD-GCN is compared against.","marker":"[27]"},{"why":"A recent high-performance denoising network used as a baseline; GD-GCN reports lower errors than it in both supervised and unsupervised settings.","marker":"[23]"},{"why":"A score-based denoising baseline used in the supervised and unsupervised comparisons.","marker":"[24]"},{"why":"The gradient-field resampling baseline that held the best sparse-cloud 1% noise result before GD-GCN.","marker":"[69]"},{"why":"The graph diffusion kernel that serves as the 'without spectral filtering' ablation baseline.","marker":"[40]"},{"why":"The multi-head attention mechanism used in Eq. (8) to approximate the Jacobian for the Riemannian metric.","marker":"[56]"}],"fun_headline_variants":["Point-cloud denoising as a neural ODE on manifolds","Micro-step graph nets evolve noisy points to clean surface","Learned Riemannian metric sharpens point-cloud denoising","Continuous-time graph convolution outperforms discrete GCNs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole 'Riemannian metric' graph construction assumes that the multi-head attention output $J_\\theta$ in Eq. (8) is a valid approximation to the true Jacobian $\\nabla z(p)$ of the learned feature map, with no error bound or dimension check; if it is not, the graph is not actually measuring Riemannian distance.","fun_headline_variants_meta":{"raw":{"variants":["Point-cloud denoising as a neural ODE on manifolds","Micro-step graph nets evolve noisy points to clean surface","Learned Riemannian metric sharpens point-cloud denoising","Continuous-time graph convolution outperforms discrete GCNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1469,"prompt_tokens":1112,"completion_tokens":357,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":728,"completion_tokens_details":{"reasoning_tokens":289}},"tokens_in":728,"tokens_out":357,"duration_ms":4422,"temperature":1.0,"reasoning_tokens":289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:28:31.035459+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train GD-GCN, then on a held-out noisy point cloud compute the finite-difference Jacobian of the map from input 3-D coordinates to the $l$-th layer feature $z(p)$ for a set of points, and compare its action on neighbor displacement vectors with the attention-based $J_\\theta$ used in Eq. (8); a systematic mismatch (wrong dimension, sign flip, or large directional error) would falsify the claim that the graph construction is geometric.","supporting_citations":[{"cited_title":"Neural ordinary differential equations,","cited_arxiv_id":null,"evidence_quote":"Supplies the neural-ODE formulation used to solve the micro-step temporal graph convolution as a continuous-time system."},{"cited_title":"BernNet: Learning arbitrary graph spectral filters via Bernstein approximation,","cited_arxiv_id":null,"evidence_quote":"Provides the Bernstein polynomial basis used to build the stable graph spectral filter."},{"cited_title":"Learning graph- convolutional representations for point cloud denoising,","cited_arxiv_id":null,"evidence_quote":"The graph-convolutional denoising baseline whose residual layer scheme MST-GConv generalizes and which GD-GCN is compared against."},{"cited_title":"Differentiable manifold reconstruction for point cloud denoising,","cited_arxiv_id":null,"evidence_quote":"A recent high-performance denoising network used as a baseline; GD-GCN reports lower errors than it in both supervised and unsupervised settings."},{"cited_title":"Score-based point cloud denoising,","cited_arxiv_id":null,"evidence_quote":"A score-based denoising baseline used in the supervised and unsupervised comparisons."},{"cited_title":"Deep point set resampling via gradient fields,","cited_arxiv_id":null,"evidence_quote":"The gradient-field resampling baseline that held the best sparse-cloud 1% noise result before GD-GCN."},{"cited_title":"Beltrami flow and neural diffusion on graphs,","cited_arxiv_id":null,"evidence_quote":"The graph diffusion kernel that serves as the 'without spectral filtering' ablation baseline."},{"cited_title":"Attention is all you need,","cited_arxiv_id":null,"evidence_quote":"The multi-head attention mechanism used in Eq. (8) to approximate the Jacobian for the Riemannian metric."}],"review_version":1}