- 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
- 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.
- 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
- 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.
- 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.
- 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.
- 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
- 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?
- 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?
- 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?
- 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
[ 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 condition | Noise std. | Base matching | AKDM + APSP | |
|---|---|---|---|---|
| Clean input | - | 88.2 / 80.3 | 90.5 / 83.2 | +2.3 / +2.9 |
| Uniform subsampling (75%) | - | 79.5 / 76.7 | 85.6 / 79.4 | +6.1 / +2.7 |
| Uniform subsampling (50%) | - | 76.2 / 67.5 | 80.1 / 69.7 | +3.9 / +2.2 |
| Non-uniform sampling (75%) | - | 77.9 / 75.8 | 85.9 / 78.6 | +8.0 / +2.8 |
| Gaussian jitter, Level 1 | 0.01r | 64.6 / 60.4 | 78.1 / 70.2 | +13.5 / +9.8 |
| Gaussian jitter, Level 3 | 0.03r | 61.8 / 54.8 | 69.4 / 62.1 | +7.6 / +7.3 |
| Gaussian jitter, Level 5 | 0.05r | 55.0 / 50.9 | 64.3 / 55.7 | +9.3 / +4.8 |
| Here, 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 (%).
| Category | Base matching | + AKDM | + AKDM + APSP |
|---|---|---|---|
| Capsule | 92.9/80.6 | 88.8/83.2 | 86.6/86.4 |
| Screen | 70.0/57.5 | 54.4/58.1 | 54.9/59.5 |
Number of anomalous objects with worsened rankings across anomaly types.
| Anomaly type | Samples per category | Capsule | Screen |
|---|---|---|---|
| Areal | 30 | 2 | 16 |
| Scratch | 30 | 0 | 3 |
| Sphere | 30 | 12 | 1 |
| Striate | 30 | 2 | 0 |
| Overall | 120 | 16 | 20 |
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
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.
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.