{"id":"0821dbf8-8906-49d7-8e07-e1043f5a5ee8","arxiv_id":"2412.16361","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Sinkhorn-regularized DRO training loss improves neural SDF reconstruction from sparse noisy point clouds compared to recent baselines.","lead":"This paper uses distributionally robust optimization to train neural signed distance functions from sparse, noisy 3D point clouds. It reports better reconstruction quality than several recent baselines on object, human, and scene datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation protocol may drive reported gains: optimal-epoch selection based on Chamfer distance to the noisy training input, plus cited baselines not run under the same protocol, threaten the central empirical claim.","rationale":"The reader's weakest assumption was the fixed λ in Eq. 11. That is a valid caveat about the DRO interpretation: without optimizing λ, Eq. 11 is a softmax-weighted loss rather than the exact dual of the Sinkhorn DRO problem, so the 'worst-case distribution' claim is softened. However, the central claim is empirical—that the framework improves reconstruction. The fixed-λ choice does not by itself make the empirical gains false; it only weakens the theoretical framing. The evaluation protocol is more directly load-bearing. Selecting the epoch by CD to the training input is a nonstandard model-selection rule, and for cited baselines (Table 3) it may not even be applied. Without error bars, the reported margins are unverified. This is a falsifiable, concrete concern: a re-run with a fixed schedule and variance estimates would either confirm or refute the headline numbers. I therefore keep the reader's CONDITIONAL verdict unchanged, but direct the condition at the evaluation protocol rather than (only) at the fixed dual variable. The paper deserves credit for honest limitation statements (Sec. 6) and for ablating ρ and λ, which suggests the authors are not hiding sensitivity; the issue is specifically the comparability of the numbers in Tables 1–3.","tokens_in":18144,"tokens_out":10740,"duration_ms":97855,"concrete_test":"Re-run SDRO, NAP, and SparseOcc on ShapeNet (same 1024-point noisy inputs) using a fixed training schedule with a pre-registered number of epochs, and report the final-epoch checkpoint; also report mean ± std over at least 5 seeds. If SDRO no longer beats both baselines on CD1/CD2/NC/FS, the central claim is not established. As a second check, select epochs using a clean validation cloud (e.g., a noiseless dense sample from the same mesh) instead of CD to the noisy input; if the ranking changes, the reported advantage is an artifact of the selection rule.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline numbers in Tables 1–4 depend on an evaluation protocol that is not held constant across all reported baselines. Section 4.3 states: \"we select the optimal evaluation epoch for each method based on Chamfer distance between the reconstructed and input point clouds.\" The input cloud is also the training signal, so this criterion rewards checkpoints that reproduce the noisy input rather than the true surface; for a method whose training curve has a different shape, the selected checkpoint can differ dramatically. Section 4.6 states that NAP and SparseOcc results in Table 3 \"are cited from their respective publications,\" so those numbers were not necessarily produced under the same epoch-selection rule. If the cited baselines used a fixed number of epochs while the proposed method is selected post hoc over training time, the reported margins (e.g., CD1 0.63 vs 0.76 in Table 1) may reflect checkpoint selection, not the DRO loss. No standard deviations or multiple seeds are reported anywhere, so the margins cannot be distinguished from run-to-run noise. The fixed-λ approximation in Eq. 11 is a real theoretical caveat, but the empirical comparison is the load-bearing part of the central claim; if the protocol artifact is real, the paper's main conclusion does not follow. The paper is otherwise clearly written and includes sensible ablations (Tables 4, 5, Figs. 9–10), which is why the concern is about evaluation rather than the method's internal logic.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a distributionally robust optimization (DRO) framework for unsupervised neural SDF reconstruction from sparse, noisy point clouds. Two variants are presented: WDRO, based on a Wasserstein uncertainty ball, and SDRO, based on a Sinkhorn distance with entropic regularization, both building on the Neural Pull baseline. The central claim is that these DRO losses improve reconstruction fidelity relative to recent baselines (NP, NAP, SparseOcc, NTPS) across object, articulated shape, and scene datasets, with qualitative and quantitative evaluations in Tables 1–4 and additional ablations in the appendix.","tokens_in":18468,"tokens_out":6802,"duration_ms":56160,"significance":"If the reported gains are reproducible, the paper makes a useful contribution by adapting tractable DRO dual formulations to SDF learning, particularly for robustness to sparse noisy inputs. The ablation studies (noise, density, hyperparameters) and training-time analysis are thoughtful and the paper is generally clearly written. However, the central empirical claim rests on an evaluation protocol that selects checkpoints by Chamfer distance to the training input, and on several baseline numbers cited from other papers; without variance estimates or controlled re-runs, the magnitude and even the direction of the improvement remain uncertain. The DRO derivation itself is standard and taken from prior work, and the paper honestly notes that its SDRO loss fixes the dual variable, which limits the formal DRO interpretation. Overall the idea is promising, but the current evidence is not strong enough to fully support the abstract's state-of-the-art claim.","major_comments":[{"comment":"The evaluation protocol selects 'the optimal evaluation epoch for each method based on Chamfer distance between the reconstructed and input point clouds.' Since the input cloud is also the training signal, this criterion selects checkpoints that reproduce the noisy input, not necessarily the true surface. For methods with different training curves, this can change the reported margins. Moreover, Section 4.6 states that NAP and SparseOcc results in Table 3 'are cited from their respective publications,' so those numbers were not necessarily produced under the same epoch-selection rule. The paper should state explicitly which baselines were re-run under this protocol, report the selected epochs, and provide multiple-seed standard deviations for at least the main tables. Without this, the headline margins (e.g., Table 1 CD1 0.63 vs. 0.76) cannot be distinguished from checkpoint-selection artifacts or run-to-run noise.","section":"Section 4.3"},{"comment":"The manuscript derives the SDRO loss from the dual of a Sinkhorn DRO problem, but the dual formulation in Eq. (9) requires optimizing the dual variable lambda. The paper fixes lambda = 20 to avoid instability. With lambda fixed, Eq. (11) is no longer the exact dual of the worst-case expected loss over a Sinkhorn ball; it is a soft-max regularized loss. The paper acknowledges this by citing [85], but it should be stated in the main text that the implemented method is a heuristic approximation and that the theoretical DRO robustness guarantees do not directly apply to the fixed-lambda objective. The claimed connection to distributionally robust optimization is therefore substantially weakened.","section":"Section 3.2 (Eq. 11), Section 4.3"},{"comment":"The text states that on Faust, 'across all metrics, our distributionally robust training procedures demonstrate superior performance.' This is directly contradicted by Table 2: NAP achieves better CD1 (0.220 vs. 0.251 for SDRO), better CD2 (0.001 vs. 0.002), better NC (0.956 vs. 0.955), and better FS (0.981 vs. 0.979). The subsequent sentence 'Notably, NAP outperforms our approach in this setting' is consistent with the table, but the preceding claim must be corrected. This overstatement affects the paper's central claim of consistent state-of-the-art improvement.","section":"Section 4.5 vs. Table 2"},{"comment":"The text claims that 'Our approach, based on Wasserstein Robust DRO (WDRO), outperforms existing methods in terms of reconstruction accuracy, as measured by CD1 and CD2.' In Table 1, WDRO has CD1 = 0.77, which is worse than NAP and SparseOcc (both 0.76), and its NC and FS are equal to or slightly worse than those baselines. Only CD2 improves (0.015 vs. 0.020). The claim should be revised to reflect that WDRO is comparable or slightly better on CD2 only, with the larger improvement coming from SDRO.","section":"Section 4.4 vs. Table 1"}],"minor_comments":[{"comment":"The symbol L is used both for the base query-pulling loss (Eq. 2) and for the combined loss on the left-hand side of Eq. (12). This is confusing; the combined loss should be denoted with a different symbol, e.g., \\mathcal{L}.","section":"Eq. (12)"},{"comment":"There is a duplicated word: 'pulling query points to their their nearest input point cloud sample.'","section":"Section 3.1"},{"comment":"The word 'aleviate' should be 'alleviate.'","section":"Section 2"},{"comment":"The text contains a typo: 'SparaseOcc' should be 'SparseOcc.'","section":"Section 4.4"},{"comment":"The paper should state explicitly which numbers in Tables 1 and 2 were produced by the authors' own runs under the optimal-epoch protocol and which were cited from the original publications; this is currently only specified for Table 3.","section":"Section 4.3"},{"comment":"The reference [46] (Liebel and Körner) is used for the loss weighting scheme, but the canonical reference for this uncertainty-based weighting is Kendall et al. (CVPR 2018). The authors should cite the original source.","section":"Eq. (12)"}],"recommendation":"major_revision","confidential_remarks":"The core idea is relevant to the journal's scope and the empirical results are suggestive, but the evaluation protocol concern is serious because it affects the validity of the main comparison. The fixed-lambda issue is a theoretical caveat that should be stated more prominently. I would encourage the authors to rerun the key baselines under the same optimal-epoch protocol, report variances over multiple seeds, and correct the text/table contradictions in Sections 4.4 and 4.5. With those changes, the paper could be acceptable; in its current form, the central claim is not yet fully supported. There is also a self-citation pattern (NAP, SparseOcc are the authors' prior work) but it does not by itself bias the empirical comparison beyond the protocol issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it is a competent application of existing DRO dual formulations to a new domain: unsupervised neural SDF fitting from sparse, noisy point clouds. The math comes from Blanchet & Murthy and Wang et al., but the combination with Neural-Pull's query distribution is new, and the authors report consistent gains over recent baselines. Second, the headline comparison is not as clean as it looks. The optimal evaluation epoch is selected per method by minimizing Chamfer distance to the training input cloud, and some baseline numbers are cited from other papers rather than rerun under the same rule. That makes the reported margins (e.g., CD1 0.63 vs 0.76 in Table 1) hard to interpret as evidence for the DRO loss specifically.\n\nWhat is genuinely good: the soft-ball worst-case query sampling is a cleaner alternative to NAP's per-point adversarial perturbations, and the Sinkhorn variant is computationally faster. The paper is clearly written, the ablations on noise, density, and hyperparameters are informative, and the limitation section is honest: NAP and SparseOcc can outperform SDRO on clean, dense inputs. The authors also cite prior DRO work properly; the fixed-λ caveat is disclosed, though not emphasized.\n\nThe main soft spot is the evaluation protocol. Selecting the epoch that minimizes Chamfer distance to the input cloud rewards reproducing the noisy input rather than the true surface, and for methods with different training dynamics this can shift the reported quality substantially. Table 3 explicitly says NAP and SparseOcc numbers are cited from their publications, so those baselines likely were not selected this way. No error bars or multiple seeds are reported anywhere, so we cannot distinguish real margins from run-to-run noise. The fixed λ=20 also means Eq. 11 is not the exact DRO dual; it is a heuristic soft-max regularizer, so the formal robustness guarantee does not carry over. That is a real caveat, but it weakens the theoretical framing rather than the method's practical plausibility.\n\nThis paper is for researchers working on unsupervised reconstruction from sparse point clouds and anyone interested in applying DRO to neural fields. It deserves a serious referee: the application is sensible, the presentation is strong, and the empirical claims are plausible but need tightening. A referee should ask for code, error bars, and a rerun of cited baselines under the same epoch-selection protocol, and should push on whether the fixed λ can be relaxed or justified. I would send it to review, not desk-reject it.","headline":"A sensible application of known DRO machinery to neural SDF fitting, but the headline empirical gains are clouded by an epoch-selection protocol that may favor the proposed method.","tokens_in":18958,"tokens_out":1849,"would_cite":true,"duration_ms":16803,"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":"Training on worst-case query distributions makes sparse 3D reconstruction robust.","keywords":["3D reconstruction","signed distance functions","sparse point clouds","distributionally robust optimization","Wasserstein distance","Sinkhorn distance","implicit neural representation","unsupervised learning"],"falsifier":"Train SDRO on the same sparse noisy benchmark while actively updating $\\lambda$ using the dual update from Eq. (6); if the actively optimized version fails to match the fixed-$\\lambda$ version, the paper's attribution of its gains to DRO is unsupported. A second check would replace the soft-max over $Q_{q,\\rho}$ with plain Gaussian noise of the same width: if reconstruction quality is unchanged, the worst-case distribution is not the operative ingredient.","tokens_in":17987,"feed_emoji":"🧊","tokens_out":8298,"duration_ms":67265,"temperature":0.7,"pith_summary":"This paper argues that a neural signed distance function (SDF) can be learned more accurately from sparse, noisy, unoriented point clouds when the network is trained on a worst-case distribution of spatial query points rather than a fixed query distribution. The authors formulate query selection as distributionally robust optimization (DRO) over a Wasserstein ball around the empirical query distribution, and make it tractable through dual reformulations. Replacing the Wasserstein distance with the entropically regularized Sinkhorn distance yields a stable and fast training objective. Across object, human, and scene benchmarks, the method improves Chamfer distance, normal consistency, and F-score over the previous state of the art, including the adversarial baseline NAP and the occupancy baseline SparseOcc.","feed_headline":"Training on worst-case queries sharpens sparse 3D shapes","feed_subtitle":"A distributionally robust loss beats prior methods on ShapeNet, Faust, and 3D Scene.","key_machinery":"The load-bearing object is the adversarially chosen query distribution $Q'$ in a Wasserstein ball around the empirical query distribution $Q$. The identity that carries the argument is the dual reformulation of the Sinkhorn DRO problem: for a fixed dual variable $\\lambda>0$, the robust loss becomes $\\lambda\\rho\\,\\mathbb{E}_{q\\sim Q}[\\log \\mathbb{E}_{q'\\sim Q_{q,\\rho}}[\\exp(L(\\theta,q')/(\\lambda\\rho))]]$, where $Q_{q,\\rho}$ is a Gaussian centered at $q$ under the quadratic transport cost $c(q',q)=\\frac{1}{2}\\|q'-q\\|^2$. This soft-max over nearby queried points spreads the SDF error across the shape instead of letting it concentrate in low-density or noisy areas.","core_discovery":"The central discovery is that the robustness problem in sparse-point-cloud SDF learning can be cast as a choice of query distribution: instead of sampling queries normally around each input point and minimizing the average pull loss, one minimizes the loss over the worst-case distribution of queries inside a Wasserstein ball around that empirical distribution. The paper shows that this DRO problem admits a practical dual formulation, and that substituting the Sinkhorn distance for the Wasserstein distance gives a smoothed worst-case distribution that is efficient to sample and backpropagate. On ShapeNet the entropic variant (SDRO) reaches a Chamfer distance of 0.63, compared with 0.76 for NAP and SparseOcc, and it also leads on the 3D Scene dataset; the authors attribute the gains to a better spatial distribution of SDF errors, which concentrate in sparse and noisy regions under the baseline.","pith_inferences":["The role of the DRO framing could be isolated by comparing SDRO against plain additive Gaussian noise on queries with the same variance: if the gains persist without the soft-max, the effect is smoothing, not worst-case hedging.","Because the paper fixes $\\lambda=20$ rather than optimizing it, the implemented method is formally a fixed soft-max regularizer; an exact or adaptive dual update might either increase robustness or reveal that the fixed choice is essential.","The same adversarial-query principle applies to other implicit-field losses, such as occupancy functions or radiance fields, wherever spatial queries are sampled during training."],"forward_implications":["If the central claim is right, a single sparse noisy scan suffices for faithful implicit reconstruction without labelled shape priors or hand-tuned smoothness weights.","The entropic SDRO variant reaches its best result in under six minutes on reported hardware, so the robustness gain does not come with a large training-time penalty.","The same unsupervised loss transfers across synthetic objects, real articulated humans, LiDAR road scenes, and multi-view stereo point clouds.","Because the method is unsupervised and architecture-agnostic, it can be dropped into existing Neural-Pull-style pipelines to upgrade their sparse-input robustness."],"supporting_citations":[{"why":"Supplies the base Neural-Pull objective, the query sampling distribution centered at input points, and the network architecture.","marker":"[54]"},{"why":"Provides the strong duality result that turns the Wasserstein DRO problem into the tractable dual used in the WDRO formulation.","marker":"[11]"},{"why":"Introduces Sinkhorn DRO and its dual, the theoretical basis for the SDRO loss and the choice of reference measures.","marker":"[85]"},{"why":"Supplies the adaptive dual-variable update rule for WDRO and the framework connecting DRO to adversarial training.","marker":"[15]"},{"why":"Defines the adversarial query baseline NAP that this paper generalizes and compares against.","marker":"[64]"},{"why":"Sparse occupancy baseline that the paper benchmarks and outperforms on sparse noisy inputs.","marker":"[67]"},{"why":"NTPS baseline; also provides the protocol for sparse point-cloud generation on the 3D Scene dataset.","marker":"[21]"},{"why":"Shows entropic regularization in Wasserstein DRO, supporting the claim that SDRO produces more diffused adversaries.","marker":"[5]"}],"fun_headline_variants":["Worst-case queries sharpen sparse 3D from noisy points","Sinkhorn-smoothed robust loss wins on sparse SDFs","Hard-query training beats baselines on sparse point clouds","Robust SDF learning via adversarial query distributions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training objective is the exact worst-case DRO loss only if the dual variable $\\lambda$ is optimized; with $\\lambda$ fixed at 20, the implemented loss is a fixed soft-max regularizer, so the formal robustness guarantees of DRO do not carry over to the algorithm that actually produces the reported results.","fun_headline_variants_meta":{"raw":{"variants":["Worst-case queries sharpen sparse 3D from noisy points","Sinkhorn-smoothed robust loss wins on sparse SDFs","Hard-query training beats baselines on sparse point clouds","Robust SDF learning via adversarial query distributions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000201,"raw_usage":{"total_tokens":1321,"prompt_tokens":831,"completion_tokens":490,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":447,"completion_tokens_details":{"reasoning_tokens":421}},"tokens_in":447,"tokens_out":490,"duration_ms":5182,"temperature":1.0,"reasoning_tokens":421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:38:25.283382+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SDRO on the same sparse noisy benchmark while actively updating $\\lambda$ using the dual update from Eq. (6); if the actively optimized version fails to match the fixed-$\\lambda$ version, the paper's attribution of its gains to DRO is unsupported. A second check would replace the soft-max over $Q_{q,\\rho}$ with plain Gaussian noise of the same width: if reconstruction quality is unchanged, the worst-case distribution is not the operative ingredient.","supporting_citations":[{"cited_title":"Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces","cited_arxiv_id":null,"evidence_quote":"Supplies the base Neural-Pull objective, the query sampling distribution centered at input points, and the network architecture."},{"cited_title":"Unsupervised occu- pancy learning from sparse point cloud","cited_arxiv_id":null,"evidence_quote":"Sparse occupancy baseline that the paper benchmarks and outperforms on sparse noisy inputs."}],"review_version":1}