acceptodds
Withdrawn from ICLR 2027

SES-3D: Beyond Spatial Correspondence for Few-Template 3D Anomaly Localization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Few-template 3D anomaly localization compares a query shape with a small set of normal references. Spatial registration restricts these comparisons to corresponding regions, but leaves a second choice: how should a handful of matched references be scored? We study this choice in a fixed registered pipeline and use Spatial Exemplar Scoring (SES-3D), the cosine distance to the nearest valid same-location normal feature. The score retains observed normal exemplars and requires no local variance estimate. With the encoder, registration, references, and point projection held fixed, it raises point AUROC from 96.36% to 98.54% and point average precision from 46.62% to 63.32% on 1,172 of 1,206 Real3D-AD inputs accepted by registration. It reaches 98.64% point AUROC on the 40 base Anomaly-ShapeNet categories. Three-template comparisons favor exemplar scoring over the raw effectsize statistic on both point metrics; two-template controls show smaller AUROC gains over mean and centroid scoring. Variance shrinkage remains competitive, and estimator preference changes with the representation. These results establish score estimation as a consequential design choice after spatial correspondence, rather than a universally optimal nearest-neighbor rule.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
3D anomaly detection, point-cloud anomaly localization, few-template learning, exemplar memory, spatial correspondence

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_VeufOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SES-3D for few-template 3D anomaly localization. It stores registered normal exemplars at each spatial location and scores a query using its nearest corresponding exemplar. Experiments on Real3D-AD, Anomaly-ShapeNet, and IEC3D-AD evaluate spatial correspondence, scoring methods, and different template budgets.

Critical strengths
  1. The WHERE–HOW evaluation clearly separates spatial correspondence from anomaly scoring.

  2. The controlled experiments use shared features, references, registration, and projection across estimators.

Critical weaknesses
  1. The methodological novelty is limited. SES-3D is essentially nearest-neighbor scoring over spatially registered features.

  2. The main SOTA comparison uses different evaluation protocols and accepted input sets across methods.

  3. The method excludes 34 Real3D samples that fail registration, which may introduce selection bias.

  4. SES-3D does not outperform the shrinkage estimator on Real3D with three templates.

  5. The reported scoring speedup changes end-to-end latency by less than 1%.

Questions for authors
  1. How would the results change if registration failures were counted as detection failures?

  2. What is the practical advantage over shrinkage scoring on Real3D?

3.Can all methods be compared using the same encoder, registration, references, and test samples?

  1. How sensitive is SES-3D to registration errors and incomplete reference coverage?

  2. Why is the scoring speedup important when it contributes less than 1% of total runtime?

Reviewer_gsijOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SES-3D studies point-level anomaly localization when only a few normal reference shapes are available. It reuses the registered patch-feature pipeline of COSSAD, including a frozen RIConv++ encoder trained with cross-category anomalous patches. At each registered location, the proposed score is the minimum cosine distance between the query feature and the valid normal reference features. Gaussian projection produces a dense point-level anomaly map; the principal object-level results use a separate inherited effect-size head.

The experimental contribution separates reference retrieval from score estimation and compares the proposed rule with raw effect-size, variance-shrinkage, centroid, and mean-distance estimators. The paper reports 98.54% point AUROC and 63.32% point AP on 1,172 successfully registered Real3D-AD test scans, and 98.64% point AUROC and 71.37% point AP on the original 40 Anomaly-ShapeNet categories. Additional experiments investigate two- and three-template budgets, alternative feature representations, and computational cost. The method introduces no additional trainable scoring parameters.

Critical strengths

S1. Substantial localization gains under a controlled scoring comparison

Table 2 holds the feature representation, registration, and projection fixed. On the accepted Real3D cohort, replacing the raw effect-size score improves point AUROC from 96.36 to 98.54 and point AP from 46.62 to 63.32. The matched ShapeNet-39 comparison also improves both metrics. These AP gains are practically meaningful for identifying small defect regions and show that the scoring component deserves attention independently of the backbone. The fixed-budget comparisons and paired category-bootstrap intervals in Table 18 provide further support; the statistical analysis appropriately uses categories rather than treating individual points as independent observations.

S2. A simple usable rule supported by informative controls

The scorer requires neither additional training nor a local variance estimate and remains defined with a single valid reference. The two-template controls in Table 11 show positive AUROC differences over both centroid and mean-distance scoring on the evaluated cohorts. The study also includes shrinkage controls, common-support comparisons for UCA, and negative transfer cases, rather than relying solely on the weakest comparator. This makes the paper useful as an empirical study of estimator choice. The reported scoring-stage speedup is an additional implementation benefit, although the paper correctly states that the end-to-end latency reduction is below 1%.

Critical weaknesses

W1. The methodological advance is narrow, and the advantage over stronger simple estimators is limited

Equation 3 is a nearest-neighbor distance restricted to a registered local memory. Nearest-neighbor normal-memory scoring is already established in PatchCore (Roth et al., 2022), while the registered local-memory pipeline is inherited from COSSAD (Tarvo et al., 2026), as acknowledged in Section 2. The centroid example and feature-space stability bound clarify the rule, but do not establish a distinctive new estimation principle. The paper therefore needs a sufficiently consequential empirical insight to justify the contribution beyond this straightforward combination.

The strongest controls temper that case. At three templates on Real3D, shrinkage achieves 98.44/63.65 AUROC/AP versus 98.41/62.98 for SES (Table 4); both paired differences are unresolved in Table 18. On ShapeNet-50, SES improves AUROC over the centroid by 0.35 percentage points, but the AP difference is only 0.45 points with a 98.75% interval of [-0.29, 1.41]. I recognize the positive ShapeNet AP result against shrinkage, but the very large headline gains are primarily against the unregularized effect-size comparator. Transfer is also representation dependent: with Point-M2AE at four templates on ShapeNet, SES loses 1.76 AUROC points and 1.78 AP points relative to effect-size scoring (Table 23).

These results establish a useful practical scoring alternative, but provide limited evidence for a broadly preferable model of local normality. In particular, Appendix A.1 defines shrinkage at fewer than three references, yet the primary two-template comparison omits this directly relevant alternative.

W2. The few-template and benchmark conclusions depend on support and coverage choices that remain insufficiently resolved

Appendix A.2, lines 655-658, constructs the shared slot set using finite features from the entire available normal bank before selecting a size-T subset. This is a sensible control for isolating scoring differences, but it does not establish what happens when only T normal shapes are available to construct the support in the first place. The omitted references can influence which slots contribute to projection. The paper should quantify whether rebuilding support from the selected subset changes coverage or performance; if the masks are identical, documenting that would resolve this concern.

The main Real3D results also cover 1,172 of 1,206 test inputs, excluding 34 registration failures, while the controlled retrieval and budget experiments use a separate 587-scan subset (Section 4.1; Appendix C). Table 1 combines these conditional results with published results under other methods' native protocols. The matched Template3D comparison is helpful, but uses one reference for Template3D and the available bank for SES. Moreover, the reference counts for the headline full-inventory Real3D results are explicitly unavailable in Table 9.

The authors disclose these limitations, and they do not negate the within-cohort improvement. Nevertheless, the available evidence does not fully establish performance under a strict reference budget or an equivalent full-benchmark protocol. This matters because few-template operation is central to the paper's claimed significance.

Questions for authors

Q1. Decision-relevant validation: how competitive is SES with shrinkage at the smallest budget?

Could you report the two-template comparison against the shrinkage estimator in Appendix A.1, using the same reference subsets, support, and projection as SES, with both point AUROC and AP and paired category intervals? Equation 11 and its accompanying conventions permit this estimator to operate below three references, so definedness alone does not distinguish SES from a regularized moment-based alternative. A consistent advantage here would strengthen the few-template motivation; comparable performance would favor a narrower interpretation centered on simplicity and avoiding the raw estimator's instability.

Q2. Clarification and validation: what information is available under the stated template budget?

For T = 2 and T = 3, can you quantify the change in slot coverage and localization metrics when support is constructed using only the selected references rather than the full bank? Please clarify whether any additional canonical reference shape is required and, if so, whether it is counted in T. Also provide the exact reference budgets for the headline Real3D results and clarify how the 587-scan cohort was selected. For the benchmark comparison, results for the strongest available baselines on the same accepted cohort, together with explicit treatment of registration failures, would help determine whether the conditional results support the broader comparison.

These requests concern validation and clarification of the current claims, rather than suggestions for a separate future research project.

Reviewer_ELWGOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SES-3D studies anomaly localization with a small number of normal reference point clouds. It builds on the COSSAD pipeline, retaining registration, frozen RIConv++ patch features, spatially indexed reference memories, and Gaussian projection to points. Its main change is to replace the effect-size localization score with the minimum cosine distance to a valid normal exemplar at the corresponding spatial slot. This scoring module adds no trainable parameters; the primary object-level output still uses the inherited effect-size branch. The paper reports localization improvements over the matched effect-size baseline on Real3D-AD and Anomaly-ShapeNet, and studies reference budgets, alternative estimators, and transfer across feature representations. The main Real3D results cover 1,172 successfully registered inputs out of 1,206 official test inputs.

Critical strengths

The substantial supplementary analyses are useful. They examine reference budgets, alternative estimators including centroid, mean-distance, and variance-shrinkage scores, transfer across representations, category-level results, and statistical intervals. These analyses help identify where the scoring change is useful and where its advantage is limited, rather than relying only on the headline comparison with raw effect-size scoring.

Critical weaknesses
  1. Limited conceptual advance is my main concern. Normal-feature memories and nearest-neighbor scoring are established in image-based PatchCore and 3D methods including Reg3D-AD, the FPFH-only baseline in BTF, and Simple3D, while registered local comparisons already appear in Template3D-AD and COSSAD. These precedents use different features, candidate sets, and scores, but together establish the main ingredients on which SES builds. SES mainly retains COSSAD's spatially aligned pipeline and replaces its effect-size localization score with the minimum cosine distance to a valid normal exemplar in the same slot, so the methodological increment appears limited rather than a new approach to normal-reference matching. The controlled study is useful, but competitive variance-shrinkage results make the evidence more persuasive as an analysis of estimator choice than as a substantial methodological advance, and I do not find the contribution sufficient for acceptance.

  2. Reproducing the complete pipeline remains difficult. I could access the repository's top-level listing but could not retrieve its source, and the available materials do not establish a complete configuration for preprocessing, registration and slot construction, checkpoint mappings, or template and test identifiers. This is particularly consequential for the main Real3D evaluation, which excludes 34 registration refusals and lacks recorded template counts in its metric export. Providing the exact upstream version and configuration, required assets, and reference/test manifests would make the reported results much easier to reproduce.

  3. The presentation could connect the proposed scoring rationale more directly to observed examples. Figure 2 explicitly states that its shapes and score illustrations are schematic; this is reasonable for a workflow diagram. Figure 3 already provides real qualitative predictions, including a failure case. A more persuasive explanation would link a real local query and its corresponding normal references to the actual effect-size and SES scores, showing how the scoring replacement changes the prediction. This is a secondary presentation suggestion, not a concern about using schematic diagrams in itself.

Questions for authors
  1. What is the main conceptual contribution beyond nearest-neighbor memory scoring applied to the spatial candidates supplied by COSSAD? Given the existing centroid, mean-distance, and shrinkage controls, could you clarify what general conclusion the study supports about when exemplar scoring is preferable, especially where shrinkage matches or exceeds it?

  2. Could you restore access to the anonymous source and identify the exact upstream version, configuration, checkpoints and category mappings, training splits, and runtime environment needed to reproduce predictions from raw point clouds? Please specify which preprocessing, registration, and slot-construction settings are inherited unchanged and which are modified.

  3. Could you provide the reference identifiers and template counts for the main Real3D evaluation, the accepted and rejected test identifiers with registration rejection criteria, and the identifiers defining the 587-scan subset? Please also identify the exact extended ShapeNet inventory. How should the reported comparisons with native-protocol baselines be interpreted given these differences in evaluated inputs?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.