{"id":"64fd6f96-f85a-4f65-b5be-fcfe1439cf75","arxiv_id":"2412.03056","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A zero-parameter point cloud classifier replaces Point-NN's sinusoidal encoding with a Gaussian encoding and reports gains, but its hyperparameters were selected using the test sets, so the gains are not independently validated.","lead":"A point cloud classifier with no trainable weights, built from farthest point sampling, neighbor search, and a fixed Gaussian encoding, reports 85.3% on ModelNet40 and 85.9% on ScanObjectNN. The result could help real-time systems, but the paper tunes its settings directly on the test sets, so the reported numbers are not unbiased estimates.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported accuracies are selected on the test sets via the Section 4.7 ablations, so the claimed margins over Point-NN are not established as unbiased estimates.","rationale":"The reader's weakest assumption identifies exactly the load-bearing flaw: Section 4.7 selects the method's hyperparameters by maximizing accuracy on the same ModelNet40 and ScanObjectNN test sets that later produce the headline numbers in Tables 1–2. This is not an internal inconsistency in the equations, but it invalidates the paper's generalization claim. The method itself is a small variant of Point-NN—Gaussian positional encoding (Eqs. 4–6) replaces sinusoidal encoding in an otherwise inherited pipeline—so the only substantive evidence for the claimed improvement is the accuracy delta. If that delta is inflated by test-set selection, the central claim collapses. A re-evaluation with a validation split and a fair Point-NN baseline is necessary before the reported margins can be accepted. I therefore agree with the reader's REJECT verdict: the current submission does not establish the stated performance, and the correct path forward is a major revision with a proper evaluation protocol. No change to the reader's verdict is needed.","tokens_in":12773,"tokens_out":3376,"duration_ms":33859,"concrete_test":"Re-run Point-GN with a proper train/validation/test protocol: split ModelNet40's official training set into a memory set and a held-out validation set (e.g., 90/10 stratified), select K, GPE dimension, number of stages, and sigma by maximizing accuracy on the validation set only, then evaluate once on the official 2,468-sample test set; repeat the same protocol on ScanObjectNN. If the validation-selected configuration yields test accuracy more than 1 point below the reported 85.29% on ModelNet40, or fails to beat a reimplemented Point-NN baseline under identical memory and evaluation code, then the Section 4.7 test-set selection is what drives the reported margin.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that zero-parameter GPE beats Point-NN by 3.5 points on ModelNet40 and by 21.5 points on ScanObjectNN PB-T50-RS—rests on hyperparameter choices made in Section 4.7. Figure 5 plots 'Accuracy (%)' on ModelNet40 and ScanObjectNN as a function of K, GPE dimension, number of stages, and sigma, with no separation between validation and test data. The text then says 'we adopted a single configuration for all three splits and aimed to find the best average performance'; these selected settings are fed directly into Tables 1–2. Because the same test sets are used both to select hyperparameters and to report final accuracy, the reported numbers are the result of test-set fitting. With zero learnable parameters, hyperparameter selection is the only form of adaptation, so this leakage is not a minor detail. The ablation itself reveals dataset-dependent optima (sigma 0.35–0.4 for ModelNet40 vs 0.3 for ScanObjectNN), and the 3.5-point and 21.5-point margins would need to survive selection on a validation split before they can be attributed to the Gaussian encoding. A secondary issue noted by the reader is that the Point-NN baseline appears lower than the original paper's reported values, which would further shrink the claimed improvement even after correcting the leakage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Point-GN, a non-parametric point cloud classification pipeline built from Farthest Point Sampling, k-Nearest Neighbors, Gaussian Positional Encoding, and a similarity-based classifier with no learnable parameters. The authors report 85.29% accuracy on ModelNet40 and an average of 85.89% across the three ScanObjectNN splits, claim improvements over the training-free Point-NN baseline, and additionally report few-shot classification results and inference speed comparisons. The central claim is that Gaussian positional encoding, combined with hand-designed feature aggregation, yields a training-free classifier that approaches fully trained models.","tokens_in":12968,"tokens_out":5301,"duration_ms":51965,"significance":"If the reported accuracies were unbiased, the contribution would be notable: a zero-parameter pipeline that comes within about two points of trained models on ScanObjectNN and improves substantially over the prior Point-NN baseline would demonstrate a useful inductive bias for point cloud classification. The paper also includes an ablation study and a public code link. However, the significance is not currently established because the hyperparameters are selected using test-set accuracy in Section 4.7, and at least one headline comparison is overstated relative to the paper's own table. The claimed margins should not be treated as externally validated results until the experimental protocol is corrected.","major_comments":[{"comment":"The ablation study reports test-set accuracy for each choice of K, GPE dimension, number of stages, and sigma, and the text states that for ScanObjectNN the authors \"adopted a single configuration for all three splits and aimed to find the best average performance.\" The same test sets are then used to produce Tables 1 and 2. Because Point-GN has zero learnable parameters, hyperparameter selection is the only form of adaptation, so the headline accuracies (85.29% and 85.89%) and the margins over Point-NN are test-selected rather than unbiased estimates. Please select hyperparameters on a validation split, freeze the configuration, and only then report accuracy on the untouched test sets.","section":"Section 4.7, Figure 5"},{"comment":"The caption claims that Point-GN \"outperforms all others,\" but the table lists PointNeXt-S at 87.7% and PointMetaBase-S at 87.9% on PB-T50-RS, both above Point-GN's 86.4%. The caption therefore contradicts the data in the same table. The sentence in Section 4.4 saying Point-GN \"consistently outperforms most\" is accurate, but the overclaim in the caption and the surrounding discussion of the gap to the best models must be corrected.","section":"Table 2 caption and Section 4.4"},{"comment":"The Point-NN baseline values (81.8% on ModelNet40 and 71.1/74.9/64.9 on ScanObjectNN) appear lower than the values reported in the original Point-NN paper for the same benchmarks. Since the claimed +3.5 and +21.5 percentage-point improvements are computed against these numbers, the authors must either reproduce Point-NN with their own code under identical conditions and report the protocol, or reconcile the discrepancy with the original publication. Without this, the relative improvement claims are not interpretable.","section":"Tables 1 and 2, Sections 4.3-4.4"}],"minor_comments":[{"comment":"The reference points v_j are used but never defined explicitly as vectors or scalars, and the concatenation notation [gamma_x, gamma_y, gamma_z]_{j=1}^V is ambiguous. Please define v_j in R^3 and write the concatenation out explicitly.","section":"Section 3.2, Eqs. (4)-(7)"},{"comment":"The symbol gamma(P_j) is used both for the retrieved features from Eq. (11) and for a new Gaussian encoding of the retrieved coordinates, which makes the update rule ambiguous. Please use distinct notation for these two quantities.","section":"Section 3.3.3, Eq. (12)"},{"comment":"The text says reference points are \"strategically chosen or learned by the model,\" but the model has no learnable parameters. This should be restated to say they are hand-set, since the claim of zero learnable parameters is central to the paper.","section":"Section 3.3.1"},{"comment":"The x-axis of Figure 5(b) is labeled only \"Dimension\" without stating whether the values are V or V times 3. The text discusses a dimension of 27, so the axis definition should be clarified.","section":"Section 4.7, Figure 5(b)"},{"comment":"The table reports mean accuracy across 10 runs, but no standard deviations are given. Please add standard deviations, especially for the 5-way 20-shot configuration where Point-GN and Point-NN are reported as tied.","section":"Table 3"},{"comment":"The inference speed comparison is shown only as a bar plot without numeric labels or error bars. A table with the measured samples per second for each dataset would make the claimed speed advantage reproducible and easier to verify.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The central obstacle is the test-set selection in Section 4.7, which invalidates the headline accuracies as unbiased estimates. If the authors can re-run the evaluation with a validation-based hyperparameter selection and report test accuracy on untouched test sets, the paper could be reconsidered. I would also ask for a direct run of the released Point-NN code to verify the baseline numbers before trusting the claimed margins."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper swaps the sinusoidal encoding in Point-NN for a Gaussian/RBF encoding, and the few-point gain it reports is not established because the hyperparameters are tuned on the test sets. The novelty is incremental but legitimate; the experimental protocol isn't.\n\nWhat's new: the method is exactly Point-NN with the positional encoding replaced by exp(-||x-v_j||^2/(2σ^2)), Eqs. 4-6, in an otherwise inherited pipeline (FPS, KNN, mean+max pooling, similarity classifier). That's a standard feature map, so the contribution is empirical rather than conceptual. The paper is clearly written, compares against the relevant non-parametric baseline, and provides a code link. If the accuracy held up, a zero-parameter classifier at 85.3% on ModelNet40 and 85.89% average on ScanObjectNN would be a useful data point.\n\nSoft spots, in order: first, the main one. Section 4.7 and Fig. 5 report accuracy on the ModelNet40 and ScanObjectNN test sets as the selection curve for K, GPE dimension, number of stages, and sigma. The text says they chose 'a single configuration for all three splits' to get the best average performance, then feed that configuration into Tables 1 and 2. With zero learnable parameters, hyperparameter choice is the only adaptation, so this is test-set fitting. The claimed margins (+3.5 and +21.5 over Point-NN) are not independent estimates. The optima are dataset-dependent (σ=0.35–0.4 for ModelNet40 vs 0.3 for ScanObjectNN), which makes the selection look even more like overfitting.\n\nSecond, Table 2's caption says Point-GN 'outperforms all others,' but PointNeXt-S (87.7) and PointMetaBase-S (87.9) beat it on PB-T50-RS. That's a factual overclaim in the caption. Third, the Point-NN baseline appears lower than the original paper's reported numbers; if that's true, the gain is inflated even before fixing the leakage. The few-shot results share the same tuning issue.\n\nWho should read it: someone working on training-free point cloud methods might want to know the idea, but only after a re-evaluation with a validation split, error bars, and reproducible code. I would not cite the current numbers.\n\nRecommendation: the empirical question is simple and checkable, so it deserves a serious referee who can require the corrected protocol. But as submitted, the headline accuracy claims are not supported.","headline":"A small, clearly-presented tweak to Point-NN, but the headline gains are undermined by hyperparameter tuning on the test sets.","tokens_in":13569,"tokens_out":6024,"would_cite":false,"duration_ms":51771,"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":"A fully non-parametric pipeline with Gaussian positional encoding classifies 3D point clouds at 85.29% on ModelNet40 and 85.89% on ScanObjectNN, with zero learnable parameters.","keywords":["point cloud classification","non-parametric network","Gaussian positional encoding","farthest point sampling","k-nearest neighbors","zero learnable parameters","ModelNet40","ScanObjectNN"],"falsifier":"Select all four hyperparameters using only a held-out validation split (or cross-validation on training data), freeze the configuration, run Point-GN once on the ModelNet40 and ScanObjectNN test splits, and compare to the reported numbers and to Point-NN under the same protocol. If the accuracies drop materially below 85.29% and 85.89%, or the margin over Point-NN collapses, the central claim is refuted.","tokens_in":12447,"feed_emoji":"🎯","tokens_out":6700,"duration_ms":60420,"temperature":0.7,"pith_summary":"This paper seeks to show that 3D point cloud classification does not need learnable parameters at all. It builds a fixed pipeline, Point-GN, that encodes each point with a Gaussian kernel against preset reference coordinates, groups points with farthest-point sampling and k-nearest neighbors, and pools features through four hierarchical stages before a similarity-based classifier compares test features to stored training features. On the standard benchmarks the authors report 85.29% accuracy on ModelNet40 and an average 85.89% on the three ScanObjectNN splits, beating the previous training-free method by 3.5 and up to 21.5 points respectively and coming within about 2 points of the best fully trained model. If true, the result matters because it would mean high-accuracy point cloud recognition is achievable with zero training, suiting real-time and resource-limited devices.","feed_headline":"Zero-parameter Point-GN hits 85.9% on real scans","feed_subtitle":"Gaussian-encoded features beat the training-free baseline by 21.5 points on ScanObjectNN and nearly match trained models.","key_machinery":"The load-bearing component is the Gaussian Positional Encoding (GPE), a fixed nonlinear map that turns raw 3D coordinates into a higher-dimensional feature vector by evaluating a Gaussian kernel centered at $V$ uniformly spaced reference points per axis, with width controlled by $\\sigma$. This encoding is applied both to the initial points and, after farthest-point sampling and k-nearest-neighbor grouping, to the gathered neighbor coordinates; the encoded neighbor features are combined with the retrieved features via $\\Gamma_j \\leftarrow \\Gamma_j + \\gamma(P_j)\\odot\\gamma(P_j)$, then pooled by mean plus max across neighbors. Four such stages are stacked, each halving the point count, and the final global feature concatenates the per-stage mean-plus-max pooled vectors. Classification is performed without learning: test features are compared by dot product against stored training features, and labels are weighted by $\\exp(-\\gamma(1-\\mathrm{Sim}))$ before a softmax, exactly in the style of Tip-Adapter. Every step is fixed, so the entire network has zero trainable parameters.","core_discovery":"On the paper's own terms, the central discovery is that replacing the usual sinusoidal positional encoding with a Gaussian positional encoding (GPE)—where each coordinate is mapped through $\\exp(-\\lVert x_i - v_j\\rVert^2/2\\sigma^2)$ for $V$ fixed reference points per axis—makes a fully non-parametric point cloud classifier accurate enough to rival trained networks. Combined with farthest point sampling, k-nearest-neighbor grouping, neighbor-mean and neighbor-max pooling, and a training-set similarity classifier with $\\exp(-\\gamma(1-\\mathrm{Sim}))$ activation, the encoder reaches 85.29% on ModelNet40 and average 85.89% on ScanObjectNN across its three splits, improving on Point-NN by 3.5 points on the synthetic set and by up to 21.5 points on the hardest real-world split. The paper also reports that in few-shot settings Point-GN exceeds or matches Point-NN in three of four configurations and beats trained parametric baselines, all with zero learnable parameters.","pith_inferences":["The same Gaussian positional encoding could be dropped into the input stage of learned 3D networks as a fixed, parameter-free augmentation, potentially improving small-data accuracy or accelerating convergence.","Memory of the similarity classifier grows linearly with the number of stored training features, so the method is most naturally suited to few-shot or prototype-based settings; scaling to very large training sets would require a nearest-centroid or hashing approximation.","The per-dataset optimal $\\sigma$ suggests a multi-scale GPE (several fixed $\\sigma$ values concatenated) might improve robustness across object scales without adding trainable parameters.","Because the entire pipeline is fixed and hand-designed, the accuracy gap to trained models is a measurable lower bound on what learned features contribute; closing that gap with learned layers could quantify the value of each component."],"forward_implications":["With zero learned weights, point cloud classifiers can come within about two points of the best trained models on real-world scans and beat the previous training-free approach by 21.5 points on the hardest ScanObjectNN split.","Deployment becomes feasible on devices without training infrastructure: inference runs at 301 samples per second on a high-end GPU, and no gradient computation or weight storage is needed.","Few-shot classification on ModelNet40 becomes nearly as strong as the non-parametric baseline, and both outperform trained parametric models, suggesting that fixed geometric features carry much of the information needed for small data regimes.","A single fixed configuration works across all three ScanObjectNN splits, removing the need for per-dataset tuning or retraining.","The Gaussian encoding outperforms sinusoidal encoding in this non-parametric setting, indicating that the choice of positional encoding function matters even without learning."],"supporting_citations":[{"why":"Provides the non-parametric framework and the training-free baseline that Point-GN extends and compares against.","marker":"[43]"},{"why":"Supplies the hierarchical FPS/k-NN local grouping architecture that Point-GN adapts.","marker":"[21]"},{"why":"Contributes the exp(-gamma(1-Sim)) activation used in the similarity-based classifier.","marker":"[41]"},{"why":"ModelNet40 benchmark that provides the synthetic classification accuracy of 85.29%.","marker":"[34]"},{"why":"ScanObjectNN benchmark that provides the real-world classification accuracy of 85.89%.","marker":"[28]"},{"why":"Inspires the mean-plus-max pooling combination used for feature aggregation.","marker":"[32]"}],"fun_headline_variants":["Zero-parameter Point-GN rivals trained networks on 3D scans","No parameters, 85.9% on real scans: Point-GN","Training-free Point-GN hits 85.9% on ScanObjectNN","Point-GN: Gaussian encoding unlocks training-free 3D accuracy","Zero-parameter classifier matches trained models on real scans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracies are only unbiased if the hyperparameters (neighbor count $K$, GPE dimension, number of stages, and $\\sigma$) were chosen without looking at the test sets; the ablation study plots test accuracy as the selection criterion, and if that choice reused the same test data, the 85.29% and 85.89% figures are not an honest measure of generalization.","fun_headline_variants_meta":{"raw":{"variants":["Zero-parameter Point-GN rivals trained networks on 3D scans","No parameters, 85.9% on real scans: Point-GN","Training-free Point-GN hits 85.9% on ScanObjectNN","Point-GN: Gaussian encoding unlocks training-free 3D accuracy","Zero-parameter classifier matches trained models on real scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001112,"raw_usage":{"total_tokens":4637,"prompt_tokens":952,"completion_tokens":3685,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":3592}},"tokens_in":568,"tokens_out":3685,"duration_ms":24583,"temperature":1.0,"reasoning_tokens":3592,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:48:44.040025+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Select all four hyperparameters using only a held-out validation split (or cross-validation on training data), freeze the configuration, run Point-GN once on the ModelNet40 and ScanObjectNN test splits, and compare to the reported numbers and to Point-NN under the same protocol. If the accuracies drop materially below 85.29% and 85.89%, or the margin over Point-NN collapses, the central claim is refuted.","supporting_citations":[{"cited_title":"Pointnet++: Deep hierarchical feature learning on point sets in a metric space","cited_arxiv_id":null,"evidence_quote":"Supplies the hierarchical FPS/k-NN local grouping architecture that Point-GN adapts."},{"cited_title":"3d shapenets: A deep representation for volumetric shapes","cited_arxiv_id":null,"evidence_quote":"ModelNet40 benchmark that provides the synthetic classification accuracy of 85.29%."},{"cited_title":"Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data","cited_arxiv_id":null,"evidence_quote":"ScanObjectNN benchmark that provides the real-world classification accuracy of 85.89%."},{"cited_title":"Dynamic graph cnn for learning on point clouds","cited_arxiv_id":null,"evidence_quote":"Inspires the mean-plus-max pooling combination used for feature aggregation."}],"review_version":1}