acceptodds
Under review as a conference paper at ICLR 2027

AMProp3D: Adaptive Matching with Probability Propagation for 3D Anomaly Detection

Abstract

Feature matching-based methods have achieved promising performance in industrial point cloud anomaly detection, yet even the same defect type can exhibit sample-dependent feature responses, while different defect patterns activate distinct latent representations and exhibit spatial continuity. Existing methods struggle to capture these variations through sample-adaptive feature matching and neighborhood-aware anomaly scoring, limiting their ability to localize diverse defects. To address this issue, we propose a sample-adaptive point cloud anomaly detection method that jointly enhances sample-relevant feature responses and local consistency of defect responses among neighboring regions. Specifically, Adaptive Kernel Distribution Matching (AKDM) estimates channel relevance from the distribution discrepancies between local test regions and normal references, and incorporates the relevance into an adaptive kernel. Maximum mean discrepancy is then employed to measure distribution-level deviations, enabling feature matching to selectively emphasize latent responses relevant to the current sample. Building on this, Anomaly Probability Surface Propagation (APSP) propagates anomaly evidence among spatially proximal and feature-consistent regions, reinforcing contiguous defect responses while suppressing isolated noise. Experiments on MiniShiftAD, Anomaly-ShapeNet, Real3D-AD, and MulSen-AD demonstrate that the proposed method improves anomaly localization across multiple competitive 3D anomaly detection baselines, validating the effectiveness of sample-adaptive feature matching and spatial consistency modeling.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
3D Anomaly Detection, Point Cloud Analysis, Feature Matching, Adaptive Kernel Learning, Distribution Matching, Anomaly Localization.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_yMnNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AMProp3D, an adaptive feature matching method for unsupervised 3D point cloud anomaly detection. Built upon FPFH geometric descriptors and a normal feature memory bank, the method introduces Adaptive Kernel Distribution Matching (AKDM), which constructs local test and normal reference feature sets, estimates adaptive channel weights from statistical discrepancies, and employs Maximum Mean Discrepancy (MMD) to compute distribution-level anomaly scores. Subsequently, Anomaly Probability Surface Propagation (APSP) combines spatial proximity and feature consistency to iteratively refine local anomaly responses, improving the spatial coherence of anomaly localization. Experiments on four benchmarks, including MiniShiftAD, Anomaly-ShapeNet, Real3D-AD, and MulSen-AD, demonstrate improvements in anomaly detection and localization performance.

Critical strengths
  1. This paper addresses the limitations of feature matching methods in adapting to diverse anomaly patterns and maintaining spatial consistency in local anomaly responses. By introducing local distribution matching and joint spatial-feature propagation, the proposed framework presents a clear motivation and offers certain practical value.
  2. The proposed method achieves competitive detection and localization performance across four 3D anomaly detection benchmarks. The main paper also provides component ablations of AKDM and APSP, comparisons of matching and propagation strategies, and sensitivity analyses of neighborhood parameters, offering empirical evidence for the effectiveness of the proposed method.
Critical weaknesses
  1. The methodological novelty of this work is relatively limited. AKDM primarily combines established MMD-based distribution measurement with Automatic Relevance Determination (ARD) feature weighting to construct an adaptive kernel based on local statistical discrepancies. APSP computes neighborhood weights using spatial distances and feature similarities, followed by iterative anomaly score propagation, which closely resembles conventional graph propagation, label propagation, and diffusion-based smoothing methods. Furthermore, the overall framework largely follows the FPFH-based feature matching paradigm of Simple3D. The two modules appear to be adaptations and combinations of existing techniques rather than fundamentally distinct methodological contributions.
  2. The experiments in the main paper do not sufficiently establish the necessity of the specific adaptive designs. Table 3 shows that, on Real3D-AD, introducing AKDM alone only improves O-AUROC/P-AUROC from 80.4%/92.3% to 80.9%/92.6%, while introducing APSP alone achieves 80.7%/92.8%. Furthermore, Section 4.4 reports that AKDM outperforms the strongest alternative matching strategy by only 0.6/0.5 percentage points, while joint spatial-feature propagation improves over spatial-only propagation by merely 0.3/0.4 percentage points. Although these results indicate that the proposed modules provide certain benefits, they do not sufficiently clarify whether the improvements primarily arise from local distribution modeling and conventional neighborhood smoothing or from the specific adaptive weighting and propagation mechanisms introduced by the authors.
  3. The underlying assumptions of AKDM and APSP are not sufficiently justified. AKDM estimates channel relevance based on mean shifts and variances between local test and normal reference features. However, larger statistical discrepancies do not necessarily indicate genuine anomalies, as they may also arise from normal geometric variations, uneven point cloud sampling densities, or scanning noise. Similarly, APSP exploits neighborhood consistency to enhance spatially coherent anomaly responses, but the propagation process may suppress genuine anomalies or spread erroneous responses when dealing with isolated defects, small-scale anomalies, or irregular boundaries. The main paper lacks targeted analyses of these scenarios, making it difficult to assess the applicability and limitations of the proposed method in complex industrial point clouds.
  4. The improvements in average performance do not fully reflect the instability across different object categories. Table 1 shows that AMProp3D improves O-AUROC/P-AUROC over Simple3D by only 0.8/0.7 percentage points on Real3D-AD. More importantly, Table 2 shows that, on MiniShiftAD, the O-AUROC for the Screen category decreases from 70.0% to 54.9%, while that for the Capsule category drops from 92.9% to 86.6%, despite the overall improvements in average performance. These results suggest that adaptive matching and probability propagation do not consistently benefit all object categories. However, the paper provides insufficient analysis of these performance degradations and the geometric conditions under which the proposed method remains effective.
Questions for authors
  1. What are the key methodological differences between AKDM and existing MMD- and ARD-based weighting methods, as well as between APSP and conventional graph propagation approaches?
  2. Section 4.4 shows that the improvements of AKDM over alternative matching strategies and joint propagation over spatial-only propagation are relatively limited. Could the authors further justify the necessity of these specific adaptive mechanisms compared with simpler alternatives?
  3. Are channel-wise statistical discrepancies sufficient to indicate anomaly relevance? How reliable are AKDM and APSP when point clouds exhibit normal geometric variations, varying sampling densities, or scanning noise?
  4. Table 2 shows noticeable O-AUROC degradations in categories such as Screen and Capsule. Could the authors analyze the causes of these performance drops and clarify under which anomaly types or geometric conditions the proposed method may fail?
3 comments

Dear reviewer:

Thank you for your thorough review and constructive feedback. We will clarify the methodological contributions and strengthen the experimental justification for our adaptive designs in the revised manuscript.

[ W1 & Q1 ] Methodological Distinctions of AKDM and APSP.

For this question, AMProp3D addresses the difficulty of identifying defect-relevant feature discrepancies and consolidating them into spatially coherent anomaly evidence. Different defect patterns can produce different responses in feature space, so fixed-metric MMD or globally learned ARD weights may not consistently emphasize the discrepancies informative for the current anomaly assessment. AKDM enables local distribution comparison to adapt to varying defect patterns, reducing the influence of less informative channels that can obscure subtle anomaly evidence under fixed-metric MMD or globally shared ARD weighting. APSP extends this adaptation to spatial refinement, where proximity alone is insufficient to distinguish coherent defect regions from nearby normal geometry. Its relevance-guided affinities encourage aggregation among spatially close regions with consistent feature responses, while bounded multiplicative residual correction preserves the original matching evidence and limits excessive smoothing. Together, the two designs support defect-sensitive anomaly measurement and spatially coherent localization.Together, these designs make both anomaly measurement and spatial refinement responsive to the current defect, helping preserve subtle discrepancies while reinforcing coherent defect regions.

[ W2 & Q2 ] Necessity of Adaptive Mechanisms.

Although the gains on Real3D-AD are relatively limited, AMProp3D consistently improves both metrics across all four benchmarks, demonstrating its broad applicability. To clarify the contributions of the proposed adaptive mechanisms beyond conventional distribution matching and spatial smoothing, we summarize the controlled ablation results on MulSen-AD below. Some results are taken from the Appendix D.

Ablation of adaptive mechanisms on MulSen-AD.

ConfigurationO-AUROC (%)P-AUROC (%)
AKDM weighting (APSP disabled)
Uniform weighting87.980.5
Mean-difference ARD88.281.8
Bayesian-inspired ARD89.182.5
APSP variants (AKDM fixed)
Without propagation89.182.5
Spatial-only propagation90.282.8
Uniform channel weights88.480.1
without residual correction84.878.9
Full APSP90.583.2

With APSP disabled, Bayesian-inspired ARD consistently outperforms uniform weighting and mean-difference ARD, demonstrating that adaptive channel relevance estimation provides additional benefits beyond local distribution matching alone.

With AKDM fixed, spatial-feature propagation outperforms spatial-only propagation. Furthermore, replacing query-adaptive channel weights with uniform weights or multiplicative residual correction with direct propagation leads to performance degradation, with the latter even underperforming AKDM alone without propagation. These results indicate that conventional neighborhood smoothing alone is insufficient and highlight the importance of adaptive affinity construction and residual score refinement. These improvements are achieved with minimal additional computational overhead. Overall, the ablations can support the effectiveness of the proposed adaptive mechanisms.

[ W3 & Q3 ] Assumptions and Robustness.

Large channel-wise discrepancies may arise from normal geometric variations or noise rather than genuine anomalies. AKDM therefore considers both distributional shifts relative to normal-reference variation and local query variation to estimate reliable channel relevance. APSP further incorporates this relevance into spatial–feature affinities to guide anomaly propagation.

We evaluate robustness on MulSen-AD using uniform and non-uniform subsampling and Gaussian jitter at severity levels 1, 3, and 5, adopting the jitter corruption method from ICML 2022. All other settings remain unchanged.

Robustness evaluation on MulSen-AD. Results are reported as O-AUROC / P-AUROC (%).

Test conditionNoise std.Base matchingAKDM + APSPΔ\Delta
Clean input-88.2 / 80.390.5 / 83.2+2.3 / +2.9
Uniform subsampling (75%)-79.5 / 76.785.6 / 79.4+6.1 / +2.7
Uniform subsampling (50%)-76.2 / 67.580.1 / 69.7+3.9 / +2.2
Non-uniform sampling (75%)-77.9 / 75.885.9 / 78.6+8.0 / +2.8
Gaussian jitter, Level 10.01r64.6 / 60.478.1 / 70.2+13.5 / +9.8
Gaussian jitter, Level 30.03r61.8 / 54.869.4 / 62.1+7.6 / +7.3
Gaussian jitter, Level 50.05r55.0 / 50.964.3 / 55.7+9.3 / +4.8
Here, rr denotes the maximum distance from the point cloud centroid.

As shown in the table, AMProp3D consistently outperforms the feature-matching baseline in both O-AUROC and P-AUROC across all tested perturbations. Although performance decreases under stronger noise, the consistent improvements demonstrate the effectiveness and robustness of adaptive distribution matching and structure-aware refinement.

In the revised manuscript, We have incorporated additional experimental results and ablation studies into appendix D.

[ W4 & Q4 ] Category- and Defect-Type Performance Analysis.

To investigate the performance degradations on Capsule and Screen, we conduct additional stage-wise and defect-type analyses, as summarized in the following two tables.

Category-level results on MiniShiftAD. Values are O-AUROC/P-AUROC (%).

CategoryBase matching+ AKDM+ AKDM + APSP
Capsule92.9/80.688.8/83.286.6/86.4
Screen70.0/57.554.4/58.154.9/59.5

Number of anomalous objects with worsened rankings across anomaly types.

Anomaly typeSamples per categoryCapsuleScreen
Areal30216
Scratch3003
Sphere30121
Striate3020
Overall1201620

While O-AUROC decreases for both categories, P-AUROC consistently improves, indicating differences between object-level detection and point-level localization. Further analysis reveals that worsened object-level rankings are mainly concentrated in sphere-type defects on Capsule and areal-type defects on Screen. These observations suggest that AMProp3D may exhibit relatively weaker performance for certain areal defects. Nevertheless, AMProp3D improves both average O-AUROC and P-AUROC on MiniShiftAD, demonstrating its overall effectiveness in detecting and localizing subtle geometric anomalies.

We have incorporated the experimental results and ablation studies into Appendix D.

Dear Reviewer yMnN,

Thank you for the time and effort you have devoted to reviewing our paper. As the discussion period is drawing to a close, we would like to kindly follow up on our detailed response to your review.We would greatly appreciate any further feedback you may have.

We appreciate your careful assessment of our work. In our rebuttal, we clarified the methodological distinctions of AKDM and APSP, added targeted ablation and robustness experiments, and provided stage-wise and defect-type analyses to explain the performance degradation on Capsule and Screen.

We hope our responses have addressed your concerns and welcome any further feedback before the discussion period ends.

Thank you again for your time and thoughtful feedback.

Best regards,

The Authors

Reviewer_VyNHOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AMProp3D, an unsupervised 3D point-cloud anomaly detection framework based on feature matching. Its first component, Adaptive Kernel Distribution Matching (AKDM), compares spatial neighborhoods of test features with query-conditioned normal references using maximum mean discrepancy. Channel weights derived from local mean shifts and variances adapt the kernel to each query. The second component, Anomaly Probability Surface Propagation (APSP), refines anomaly scores through iterative propagation using spatial proximity and feature similarity, followed by a residual score correction. Experiments on four benchmarks show improved performance.

Critical strengths
  1. The method improves average object-level and point-level AUROC on all four datasets. The method also offers good accuracy–efficiency trade-off.
  2. The ablation study supports the complementary roles of AKDM and APSP.
Critical weaknesses

The novelty appears limited. APSP follows a standard graph-propagation formulation, while the benefits of its adaptive affinities and residual score correction over conventional alternatives are not sufficiently established.

Questions for authors

Can you provide two ablations to clarify the contributions of APSP: (1) replace the query-adaptive channel weights with uniform weights when constructing the spatial–feature affinities, and (2) use the propagated scores directly, mapped back to the original score scale, instead of applying the multiplicative residual correction?

Can you use \citep{} for parenthetical citations?

3 comments

Dear reviewer:

Thank you for reviewing our response. We greatly appreciate your valuable feedback. In the revised manuscript, we have revised the citation formatting.

[W1] Novelty and Effectiveness of APSP.

The central contribution of AMProp3D is to use local test-normal distribution discrepancies to adapt both anomaly matching and evidence propagation to each query region. Compared with fixed-metric nearest-neighbor scoring, AKDM estimates channel relevance from query-conditioned feature distributions and incorporates it into an adaptive kernel for MMD-based anomaly scoring. APSP carries this relevance into spatial-feature affinities, allowing anomaly evidence to propagate according to the feature channels relevant to the current query. This establishes a direct connection between the criterion used to detect deviations and the criterion used to aggregate neighboring evidence. Finally, the multiplicative residual correction uses propagation-induced changes to refine the original matching scores while retaining their scale. Together, these designs address sample-dependent geometric deviations and spatially coherent defects within one framework.

[Q1] Ablation of Query-Adaptive Channel Weighting.

To isolate the contribution of query-adaptive channel weighting in APSP, we replace the adaptive weights with uniform weights only when constructing the spatial-feature affinities, while keeping AKDM and all other experimental settings unchanged.

Comparison of query-adaptive channel weights and uniform weights.

DatasetQuery-adaptive weightsUniform weightsAKDM only
MulSen-AD90.5 / 83.288.4 / 80.189.1 / 82.5
Real3D-AD81.2 / 93.079.4 / 85.780.9 / 92.6

As shown in Table, query-adaptive weighting outperforms uniform weighting on both datasets. Notably, on Real3D-AD, uniform weighting yields lower performance than both query-adaptive weighting and AKDM alone without APSP propagation. A similar trend is observed on MulSen-AD.

These results indicate that propagation alone does not necessarily improve performance: propagation with uniform channel weights can underperform the no-propagation configuration, whereas query-adaptive weighting yields performance gains. These results highlight the importance of incorporating query-adaptive channel weights into APSP's affinity construction to guide anomaly evidence propagation.

We have incorporated additional experimental results and ablation studies into Section 4.Experiments along with discussions of the finding.

[Q2] Ablation of Multiplicative Residual Correction.

To isolate the contribution of multiplicative residual correction in APSP, we replace it with direct propagation, using the propagated scores mapped back to the original score scale as the final anomaly scores, while keeping AKDM and all other experimental settings unchanged.

Comparison of multiplicative correction and direct propagation.

DatsetMultiplicative correctionDirect propagationAKDM only
Mulsen-AD90.5/83.284.8/78.989.1/82.5
Real3D-AD81.2/93.080.6/89.380.9/92.6

In the table, multiplicative residual correction outperforms direct propagation on MulSen-AD and Real3D-AD in both object-level detection and point-level localization. Moreover, direct propagation performs worse than AKDM alone without propagation, whereas multiplicative residual correction further improves both metrics.

Together with the channel-weighting ablation, these results demonstrate the importance of controlling both evidence aggregation and score refinement in APSP. Query-adaptive channel weights guide the aggregation of neighboring anomaly evidence, while multiplicative residual correction incorporates propagation-induced changes as bounded relative adjustments to the original AKDM scores. Under the evaluated settings, these two complementary designs enable more effective refinement of the initial matching evidence than uniform weighting or direct score replacement.

In the revised manuscript, We have incorporated additional experimental results and ablation studies into appendix D.

Dear Authors,

Thank you for your detailed response and additional ablation studies. The results help clarify the contributions of APSP. I believe my main concerns have been adequately addressed, and I will maintain my original positive assessment of the paper.

I noticed that a \citep{} formatting issue remains at line 942. Please also check the supplementary material for similar minor formatting inconsistencies.

Best regards,

Reviewer VyNH

Dear reviewer VyNH,

We greatly appreciate your encouraging assessment and your recognition of the improvements in our work.Following your suggestion, we have carefully reviewed the citation formatting again throughout the manuscript and supplementary materials. Your thoughtful feedback has helped us strengthen the presentation and overall rigor of our manuscript.

Best regards,

The Authors

Reviewer_CDBkOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AMProp3D addresses point-cloud anomaly detection and localization using normal training samples. It extracts multi-scale FPFH descriptors, aggregates them spatially into regional features, and constructs a normal memory bank for nearest-neighbor anomaly scoring. Adaptive Kernel Distribution Matching (AKDM) compares local test features with retrieved normal references, derives channel weights from their mean and variance statistics, and combines an ARD-kernel MMD score with the baseline score. Anomaly Probability Surface Propagation (APSP) then propagates anomaly evidence using spatial proximity and feature similarity, retains the initial evidence through a fidelity term, and uses the propagation-induced change to refine the anomaly scores. Experiments on MiniShiftAD, Anomaly-ShapeNet, Real3D-AD, and MulSen-AD report improvements in object-level detection and point-level localization over the underlying baseline.

Critical strengths
  1. Table 3 reports improvements over the underlying baseline in both O-AUROC and P-AUROC on all four benchmarks. Testing AKDM and APSP separately and together helps identify each module’s contribution, with the combined method giving the highest scores in these ablations. This provides useful evidence that both additions contribute to the reported gains across different datasets.

  2. The added inference cost is small under the authors’ measured setting. On MulSen-AD, Appendix Table 12 reports an increase from 88.2/80.3 to 90.5/83.2 in O-AUROC/P-AUROC, while mean inference latency rises from 970.54 ms to 982.12 ms, an additional 11.58 ms (1.193%). This is a favorable improvement relative to the baseline’s inference cost in that setting.

Critical weaknesses

The experiments include recent baselines such as Simple3D and HPPF, but leave out several relevant published methods: CASL is cited in the related work without an experimental comparison, while AARD (Geometry-Aligned and Anomaly-Aware Reconstruction for 3D Anomaly Detection) and IRM3D-AD (An Information-aware Reconstruction Model for Multi-category 3D Anomaly Detection) are not covered. Since these methods address the same general point-cloud anomaly detection problem, their inclusion would help assess AMProp3D against representative approaches beyond the current baseline set. This gap makes it difficult to judge how broadly the reported performance advantages hold. Comparable results, together with a clear account of differences in pretraining, training categories, and evaluation protocols, would strengthen the evaluation; where a direct comparison is not possible, the claims should reflect that limitation.

Questions for authors
  1. Could you provide comparisons with CASL, AARD, and IRM3D-AD on shared benchmarks where the settings are comparable, or explain the specific obstacles to doing so? In particular, CASL is already cited but is absent from the experimental comparisons.

  2. How do these methods differ from AMProp3D in pretraining data, training categories, and evaluation protocols, including the categories used to compute benchmark averages? Please make these differences explicit and clarify which performance claims remain supported under comparable settings.

3 comments

Dear reviewer:

Thank you for reviewing our rebuttal. We sincerely appreciate your thoughtful and valuable feedback.

[ W1 & Q1 ] Comparisons with Recent Methods.

We reviewed the original papers of CASL, AARD, and IRM3D-AD. Since AARD and IRM3D-AD are not open source, we were able to reproduce only CASL. The results for AARD and IRM3D-AD were obtained from their original papers and supplementary materials. And we evaluated CASL on MulSen-AD using checkpoints pre-trained on Real3D-AD and Anomaly-ShapeNet due to the limited time. The quantitative comparison results are presented in the following table.

Comparison of four methods on three 3D datasets. Results are reported as O-AUROC/P-AUROC (%). "-" indicates that the corresponding result is unavailable. * Indicates that the value is calculated by averaging over 15 macro-categories, whereas the other methods use 40 fine-grained subcategories.

methodReal3D-ADAnomaly-ShapeNetMulSen-AD(4-shot)MulSen-AD
AARD82.0/86.093.1*/93.6-/--/-
IRM3D-AD81.5/83.789.9/-71.2/--/-
CASL82.3/88.288.3/90.2-/-69.6/79.2
AMprop3D81.2/93.0(88.5/88.7*)/93.980.3/78.690.5/83.2

We have incorporated the reproduced CASL results into Section 4 (Experiments) of the revised manuscript, provided additional comparisons in Appendix E, and included references to AARD and IRM3D-AD in the Related Work section.

[ Q2 ] Experimental protocols and performance comparability.

To clarify the differences in experimental settings, we examine the pretraining data, training categories, evaluation protocols, and category coverage of the compared methods. The results are shown in the following table.

Comparison of Experimental Protocols. The numbers under the dataset represent the categories used. "*" AARD uses 15 macro-categories for O-AUROC and 40 fine-grained categories for P-AUROC. “--” indicates unused or unreported information.

MethodPretraining DataTraining StrategyAnomaly-ShapeNetReal3D-ADMulSen-AD
AMProp3D (Ours)--Category-specific memory bank; training-free401215 (All normal samples)
AARD (CVPR 2026)--Diffusion reconstruction with synthetic anomalies15*/4012-
IRM3D-AD (IEEE TAI 2026)PointMAE (ModelNet40-8K)Unified multi-category reconstruction401215 (4-shot)
CASL (AAAI 2026)Normal training samples from 52 categories across Real3D-AD and Anomaly-ShapeNetJoint self-supervised pretraining; category-specific fine-tuning4012-

Comparisons with Reconstruction- and Learning-Based Methods.

The table highlights the key differences among these methods. AMProp3D is training-free, whereas AARD requires diffusion reconstruction training, IRM3D-AD uses pretrained PointMAE for multi-category reconstruction, and CASL adopts cross-dataset pretraining followed by category-specific fine-tuning.

Unlike AARD, IRM3D-AD, and CASL, which require neural network training, AMProp3D is training-free. On Real3D-AD, AMProp3D achieves 93.0% P-AUROC, outperforming AARD (86.0%), IRM3D-AD (83.7%), and CASL (88.2%) in point-level localization while maintaining competitive object-level detection performance.

On Anomaly-ShapeNet, AARD reports O-AUROC over 15 macro-categories and P-AUROC over 40 subcategories. We additionally provide 15-category averages (marked with *) for fair comparison. AMProp3D surpasses CASL and AARD in P-AUROC but trails AARD in O-AUROC, while maintaining competitive overall performance.

On MulSen-AD, we additionally evaluate AMProp3D under the same 4-shot setting as IRM3D-AD. AMProp3D outperforms IRM3D-AD by approximately 9 percentage points in O-AUROC and substantially surpasses our reproduced CASL results, including an approximately 4-point gain in P-AUROC. These results demonstrate that our method remains effective even in few-shot scenarios.

On the recently introduced MiniShiftAD benchmark, AMProp3D achieves 70.0% O-AUROC and 69.8% P-AUROC, outperforming the feature-matching baselin,respectively, demonstrating its ability to detect and localize subtle defects in high-resolution point clouds.

Overall, AMProp3D outperforms the compared recent learning- and reconstruction-based methods in point-level anomaly localization, while maintaining competitive object-level detection-all without external pretraining or neural network optimization. Controlled experiments across four benchmarks validate the effectiveness of AKDM and APSP. Together with its minimal computational overhead, AMProp3D achieves a favorable balance among detection accuracy, localization quality, and efficiency, particularly for subtle geometric defects. We have added the detailed protocol analysis in appendix E.

Thank you for the authors’ response. Most of my concerns have been addressed, and I have raised my rating to Weak Accept.

We sincerely thank you for your constructive feedback and for recognizing the improvements in our work.Your comments have been invaluable in helping us improve both the clarity and the quality of our work.

Comments on the paper

1 comment

Dear Reviewers,

A friendly reminder to check the authors’ rebuttals and engage in the discussion, as the deadline is approaching. Authors still have time to address any further questions or concerns you may have, so your timely feedback would be greatly appreciated.

Thank you for your time and effort!

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.