{"id":"6f25d211-aebd-4d8a-b67c-0b4ef803051b","arxiv_id":"2412.13461","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 3D anomaly detection method that uses internal z-axis projection slices and Laplacian feature filtering reports state-of-the-art results on Real3D-AD and Anomaly-ShapeNet.","lead":"Researchers built a 3D anomaly detector that slices point clouds horizontally and looks at the inside surfaces, not just the outside. It reports large pixel-level gains on the Real3D-AD benchmark, but the paper has inconsistencies, no code, and parameters tuned on the test data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The attribution of the SOTA gains to internal slices is unsupported: Eq. (5) is tautological, and the ablation is confounded by test-benchmark tuning and an unresolved 0.757-vs-0.767 discrepancy.","rationale":"The reader's weakest-assumption analysis correctly identifies that Eq. (5) is a tautology, not a proof of discriminative utility. I agree with that assessment and consider it a genuine weakness of the paper's motivation and of the claim that SIE's generalization ability is 'theoretically proven.' However, the more load-bearing issue for the empirical SOTA claim is that the only direct ablation for the internal slices is entangled with feature-filtering parameters selected on the same benchmark, and the reported O-AUROC disagrees between the text and Table 1. These issues do not prove the method is wrong, but they make the central quantitative claim unverifiable from the manuscript alone. Since the paper already receives a CONDITIONAL verdict, I do not propose moving to REJECT or UNVERDICTED; instead, the identified concerns sharpen the conditions under which the claim can be accepted: code or a controlled re-run must resolve the numeric discrepancy and show that the internal-slice gain survives untuned hyperparameters. The agreement is partial because the reader focuses on the tautology itself, while I take the tautology plus the benchmark-tuned ablation and the numeric inconsistency together as the decisive obstacle to accepting the SOTA attribution.","tokens_in":13398,"tokens_out":7852,"duration_ms":69957,"concrete_test":"Obtain the official implementation or reimplement Algorithm 1 and run Real3D-AD under a pre-registered protocol: fix alpha, beta, and gamma on a held-out category or leave them at untuned defaults, then recompute the per-category means from Table 1 to determine whether the true O-AUROC is 0.757 or 0.767. Then re-run ISMP versus ISMPO (external-only projections) with identical hyperparameters and multiple random seeds. If the corrected O-AUROC is 0.757, or if the ISMP minus ISMPO gap is within seed noise, the central internal-modality attribution loses its quantitative support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ISMP reaches SOTA by exploiting internal spatial pseudo-modalities through SIE. For that attribution to hold, the internal slices (P2, P3) must do real work, and the reported numbers must be trustworthy. The paper's theoretical support is Eq. (5): I_global = I_top + sum_{i: z_i >= z_mid}(z_i - z_mid) >= I_top. This only says that adding non-negative numbers to I_top does not decrease it; it says nothing about whether the added terms are discriminative for anomalies. The only direct evidence for the internal modality is the ablation in Table 3, but the feature-filtering hyperparameters (alpha=0.2, beta=0.2, gamma=0.001) are chosen in the 'Analysis of the Feature Filtering Module' by inspecting mean/variance heatmaps on Real3D-AD features (Fig. 4), i.e., on the same benchmark used for the final SOTA numbers. With four normal training samples per class, selection on test-set statistics can easily inflate AUROC. In addition, O-AUROC is reported as 0.757 in the main text, matching the abstract's 3.2% improvement, while Table 1's mean is 0.767, corresponding to a 4.2% improvement. These two problems together mean the SOTA claim is not yet robustly established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ISMP (Internal Spatial Modality Perception) for 3D anomaly detection. The method's central component, the Spatial Insight Engine (SIE), projects a registered point cloud into four slices (P1-P4) and uses internal slices as pseudo-modalities; an enhanced feature extraction branch combines PointMAE patch features with FPFH descriptors, and a graph-Laplacian-based feature filtering module suppresses redundant information. Anomaly scores are computed from nearest-neighbor distances in feature and coordinate memory banks. On Real3D-AD the paper reports O-AUROC 0.767 and P-AUROC 0.836 in Table 1 (with the abstract claiming 3.2% and 13.1% improvements), P-AUROC 0.691 on Anomaly-ShapeNet in Table 2, and auxiliary classification and segmentation experiments for SIE on ModelNet40 and ShapeNet-Part. The authors argue that internal information is richer than external projection and state that the generalization ability of SIE has been theoretically proven.","tokens_in":13743,"tokens_out":7020,"duration_ms":63535,"significance":"If the empirical results hold, the paper makes a useful contribution: it demonstrates that internal projection slices can serve as pseudo-modalities for 3D anomaly detection, with particularly strong pixel-level localization gains on Real3D-AD. The ablation study (Table 3) and the generalization experiments (Tables 5 and 6) provide practical evidence that the modular components, including SIE, are not obviously harmful and may help on standard benchmarks. The use of public benchmarks is another strength. However, the analytical support for the central novelty is currently only an inequality that holds by construction, the headline O-AUROC numbers are mutually inconsistent, and the parameter-selection protocol does not fully exclude test-set influence. These issues need to be resolved before the state-of-the-art claim is fully credible.","major_comments":[{"comment":"The derivation in Eqs. (1)-(5) shows only that I_global = I_top + sum_{i: z_i >= z_mid}(z_i - z_mid) >= I_top, which is true by construction because all added terms are non-negative. This does not establish that the internal z-discrepancy carries anomaly-discriminative information, and Eq. (6) is a heuristic threshold rule rather than a proof. Since SIE is the paper's central novelty and the abstract claims that its generalization ability has been theoretically proven, this is a load-bearing gap: please either provide a genuine argument (e.g., under stated assumptions, internal slices separate anomalous from normal points), or remove/reword the theoretical claims.","section":"Method / Spatial Insight Engine, Eq. (5)"},{"comment":"The reported O-AUROC numbers are mutually inconsistent. The abstract claims a 3.2% improvement, the main text reports ISMP with O-AUROC 0.757, and Table 1(a) lists a mean O-AUROC of 0.767 for ISMP versus 0.725 for IMRNet, which corresponds to a 4.2 percentage-point improvement. Please correct the numbers and state clearly which mean (0.757 or 0.767) is the official result and against which baseline the improvement is measured.","section":"Abstract and Experiments / Main Results vs Table 1"},{"comment":"The hyperparameters alpha=0.2, beta=0.2, gamma=0.001 for the feature filtering module are justified by inspecting mean/variance heatmaps of feature matrices (Figure 4), and the same parameter values are then used to obtain the final Real3D-AD results. The text suggests the heatmaps are computed on synthetic feature matrices resembling PointMAE features, but no separate validation split is used for parameter selection, and each category has only four normal training samples. This creates a risk of inadvertent test-set tuning that can inflate the reported gains and confounds the attribution of improvements to SIE and the filter. Please justify parameter selection on a held-out validation protocol, report sensitivity of the main results and the ablation to these parameters, and provide variance or confidence intervals over multiple runs or splits.","section":"Experiments / Ablation Study and Figure 4"},{"comment":"The abstract states that the strong generalization ability of SIE has been theoretically proven, but no theorem, proof, or formal statement appears anywhere in the manuscript; Tables 5 and 6 provide empirical demonstrations only. Please either supply the proof or remove the claim, and if the proof is deferred to a supplementary document, cite it explicitly.","section":"Abstract and Evaluation of SIE Generalization"}],"minor_comments":[{"comment":"The typeset formula for s_F appears to be missing a closing parenthesis or brace; please rewrite it unambiguously and define all symbols, including P(x_test) and N_3(m*).","section":"Method / Anomaly Score Calculation, Eq. (13)"},{"comment":"The patch radius r in Eq. (8) and the number of FPS center points m are never specified in the experimental details; please state their values and the patch size used.","section":"Method / Enhanced Feature Extraction"},{"comment":"Lines 16-17 normalize the filtered feature matrix by its maximum; please clarify whether this normalization is applied per test sample at inference time and justify its effect on nearest-neighbor distances in the memory banks.","section":"Method / Feature Filtering Module, Algorithm 1"},{"comment":"The text refers to the Laplacian transform, but Eq. (9) defines the graph Laplacian matrix L = D - A; please align the terminology with the actual linear-algebraic object being used.","section":"Method / Feature Filtering Module"},{"comment":"Figure 2 is difficult to read: the projection slices P1-P4 and the flow through the SIE, EfficientNet, and the memory banks are not clearly labeled; Figure 4's axes labels are also ambiguous. Please provide a higher-resolution figure with explicit axis labels.","section":"Figure 2 and Figure 4"},{"comment":"Table 2's 40 category names are concatenated into very wide rows, making it difficult to map results to categories; please reformat, for example by transposing the table or splitting it into two parts.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the main risk is not the absence of a theorem per se, but the mismatch between the claimed SOTA improvements and the reported numbers, and the lack of a clear protocol for parameter selection with only four training samples. If the authors can provide a corrected table, a proper ablation with variance, and an honest framing of Eq. (5) as motivation rather than proof, the paper could be publishable after major revision. I would also ask the editor to verify the relation of this work to prior and concurrent work on internal views of point clouds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the internal-slicing idea is genuinely new for 3D anomaly detection, and the ablation suggests it does real work. But the paper undercuts itself with two different O-AUROC numbers in the abstract, main text, and Table 1, and the 'theoretical proof' of SIE's superiority is a tautology. Worth refereeing, but not as-is.\n\nWhat's actually new: instead of projecting the point cloud from the outside, SIE slices it along the z-axis into top, bottom, and two internal partitions, then feeds those pseudo-images into an EfficientNet backbone. That's a simple, transferable idea. The local features combine PointMAE patches with FPFH descriptors around FPS centers, and the Laplacian filtering is a reasonable denoising step. The ablation in Table 3 is the most informative part: removing the internal slices drops O-AUROC by roughly 10 points, which is the strongest evidence that the internal signal matters. The generalization experiments on ModelNet40 and ShapeNet-Part are a nice extra, even if the gains are small.\n\nSoft spots, in order of seriousness. First, the numbers don't match: the abstract says 3.2% O-AUROC improvement, the experimental section says 0.757, and Table 1's mean is 0.767. That 0.757-vs-0.767 discrepancy is exactly the difference between a 3.2% and a 4.2% gain, so it's not cosmetic. Second, Eq. (5) is just saying that adding non-negative terms to a sum doesn't make it smaller. It does not prove that internal slices carry discriminative anomaly information. The ablation does that job, but the paper used Eq. (5) as if it were a proof, which is misleading. Third, the hyperparameters α, β, γ are chosen by inspecting feature-mean/variance heatmaps on Real3D-AD itself. With only four normal training samples per category, that's effectively tuning on the target benchmark. No error bars, no code, no held-out validation protocol. These problems don't sink the empirical claim, but they do mean the SOTA claim is not yet robustly established.\n\nWho this is for: the 3D anomaly detection community, specifically people working on Real3D-AD and pseudo-modal approaches. The paper is clearly written and the method is easy to describe, so a serious referee could get through it quickly. The verdict should be 'revise' rather than 'reject', with the numerical inconsistency and the hyperparameter selection protocol as the main requirements.\n\nRecommendation: send it to peer review, but require the authors to reconcile the reported numbers, show error bars or a justification for the discrepancy, and either move hyperparameter selection to a validation split or otherwise show it isn't inflating the results.","headline":"A simple internal-slicing idea shows real gains on Real3D-AD, but sloppy reporting and a tautological proof need fixing before the SOTA claim is credible.","tokens_in":14244,"tokens_out":2601,"would_cite":false,"duration_ms":24502,"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":"Looking inside point clouds, not just at their surface, yields state-of-the-art 3D anomaly detection on Real3D-AD.","keywords":["3D anomaly detection","point cloud","internal spatial pseudo-modality","Spatial Insight Engine","Real3D-AD","anomaly localization","feature filtering","AUROC"],"falsifier":"Train ISMP on Real3D-AD with the two internal slices P2 and P3 replaced by slices at randomly chosen z-depths (for example, at 30% and 70% of the height range), keeping all other components fixed. If the randomized-slice variant matches or exceeds the original P-AUROC of 0.836, then the specific mid-height partition is not the source of the improvement, and the claimed 'internal information' is an artifact of having more projections.","tokens_in":13214,"feed_emoji":"🔍","tokens_out":8523,"duration_ms":63825,"temperature":0.7,"pith_summary":"This paper argues that 3D anomaly detection methods that only view the external surface of a point cloud are missing discriminative internal structure. It introduces ISMP, which uses a Spatial Insight Engine to project the point cloud into four internal spatial pseudo-modalities, and combines them with enhanced local features and a Laplacian-based feature filter. On the Real3D-AD benchmark, ISMP raises object-level AUROC by 3.2 points and pixel-level AUROC by 13.1 points over prior state-of-the-art. The central claim is that internal views carry information about defects that external views cannot reveal, and the experiments support this by showing the internal slices alone outperform an external-only variant.","feed_headline":"Looking inside point clouds sets a new bar for 3D anomaly detection","feed_subtitle":"Internal spatial pseudo-modalities lift pixel-level AUROC by 13.1 points on Real3D-AD.","key_machinery":"The Spatial Insight Engine (SIE) is the load-bearing component. It converts a registered point cloud into four 2D pseudo-images: one top-down projection and three internal slices that bisect the cloud along the z-axis, encoding depth from the midpoint toward the top or bottom. These pseudo-images are fed to a pre-trained image encoder (EfficientNet) to produce global features, which are then aligned with local patch features from Farthest Point Sampling, PointMAE, and FPFH descriptors, and finally passed through a Laplacian-based feature filtering module that suppresses redundant information. The resulting feature matrix is compared against memory banks of normal training samples to compute point-level anomaly scores.","core_discovery":"The paper's central discovery is that internal spatial pseudo-modalities, generated by slicing a point cloud at its top, two middle planes, and bottom, provide a richer global feature representation for anomaly detection than external-only projections. The authors formalize this with an information measure $I_{\\text{global}} = I_{\\text{top}} + \\sum_{i: z_i \\geq z_{\\text{mid}}} (z_i - z_{\\text{mid}}) \\geq I_{\\text{top}}$, where the internal slices add non-negative depth-based terms to the top-down projection. Empirically, using only the two internal middle slices already outperforms an external-only variant, and the full SIE with all four slices yields the best accuracy, supporting the claim that looking inside contributes genuine signal. The method also demonstrates that the internal pseudo-modality generalizes to point cloud classification and segmentation, where injecting SIE features into a standard point cloud backbone yields small but consistent gains.","pith_inferences":["The paper's information-theoretic justification ($I_{\\text{global}} \\geq I_{\\text{top}}$) is a tautology because it follows from summing non-negative terms; it does not prove the internal depth terms are discriminative, so the empirical gains are the real evidence.","A testable extension is to apply SIE to symmetric or textureless objects where internal geometry is the only cue, to see if internal pseudo-modalities alone can detect subtle deformations.","The idea of 'internal slicing' could be transferred to voxel grids or meshes, potentially improving other 3D representation learning tasks.","Ablating the image encoder to a lighter network could reduce the inference overhead while preserving most of the benefit, as the current FPS numbers show SIE is slower than some baselines."],"forward_implications":["Future 3D anomaly detection can be built around single-sensor point clouds without needing aligned RGB-D or multi-view data, since internal views come from the same point cloud.","The SIE's generalization to classification and segmentation suggests the internal pseudo-modality is a reusable representation for point cloud understanding beyond defect detection.","The large pixel-level AUROC gain indicates that internal slices help localize anomalies, which could directly aid industrial inspection and robotic quality control.","The method operates from as few as four normal training samples per category, making it applicable to production lines where defect data is scarce.","The feature filtering module's controllability over feature mean and variance gives practitioners a principled way to tune detection sensitivity."],"supporting_citations":[{"why":"Supplies the Real3D-AD benchmark dataset and the evaluation protocol used for the main results.","marker":"Liu et al. 2023"},{"why":"Provides the complementary pseudo-modal baseline (CPMF) that ISMP extends by adding internal views.","marker":"Cao, Xu, and Shen 2023"},{"why":"Supplies the PointMAE patch feature extractor used for local features around sampled center points.","marker":"Pang et al. 2022"},{"why":"Provides the FPFH geometric descriptor used to enhance local feature representation.","marker":"Rusu, Blodow, and Beetz 2009"},{"why":"Supplies the EfficientNet image encoder that processes the SIE projection slices into global features.","marker":"Tan and Le 2020"},{"why":"Defines the classical feature baseline (BTF) that ISMP compares against.","marker":"Horwitz and Hoshen 2022"}],"fun_headline_variants":["Internal views of point clouds cut anomaly detection errors by 13.1%","Peeking inside point clouds improves 3D anomaly detection","Slicing point clouds internally lifts anomaly detection AUROC","Look inside point clouds: 13.1% pixel-level AUROC gain","Internal pseudo-modalities sharpen 3D anomaly detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The assumption that the internal slice features are actually discriminative for anomalies, rather than merely adding non-negative numbers to an information measure, is the point on which the method's motivation rests.","fun_headline_variants_meta":{"raw":{"variants":["Internal views of point clouds cut anomaly detection errors by 13.1%","Peeking inside point clouds improves 3D anomaly detection","Slicing point clouds internally lifts anomaly detection AUROC","Look inside point clouds: 13.1% pixel-level AUROC gain","Internal pseudo-modalities sharpen 3D anomaly detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00088,"raw_usage":{"total_tokens":3807,"prompt_tokens":952,"completion_tokens":2855,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":2767}},"tokens_in":568,"tokens_out":2855,"duration_ms":20418,"temperature":1.0,"reasoning_tokens":2767,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:06:12.082765+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train ISMP on Real3D-AD with the two internal slices P2 and P3 replaced by slices at randomly chosen z-depths (for example, at 30% and 70% of the height range), keeping all other components fixed. If the randomized-slice variant matches or exceeds the original P-AUROC of 0.836, then the specific mid-height partition is not the source of the improvement, and the claimed 'internal information' is an artifact of having more projections.","supporting_citations":[{"cited_title":"B.; Blodow, N.; and Beetz, M","cited_arxiv_id":null,"evidence_quote":"Provides the FPFH geometric descriptor used to enhance local feature representation."}],"review_version":1}