acceptodds
Withdrawn from ICLR 2027

Back to Memory: Local Normal Support and Reference Routing for 3D Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Memory-bank methods for 3D anomaly detection must select appropriate normal references and distinguish defects from allowable local variation. Feature-only retrieval can match defects to similar geometry in incompatible regions, while nearest-neighbor distances do not explicitly describe the variation permitted by each normal pattern. We propose BTM, a training-free framework that separates normal memory construction, reference routing, and anomaly scoring. Manifold-Aware Memory builds local normal patterns from graph-geodesic neighborhoods of fixed FPFH features, storing local subspaces, finite tangential supports, and orthogonal thicknesses. Spatial-Context Gated Routing restricts candidates by registered position and selectively reranks low-similarity queries using multiscale context. Corrected Tangent-Normal Scoring measures only deviations beyond the selected patterns' support bounds. BTM requires only normal training point clouds, without learned encoders or synthetic anomalies. It achieves P-AUROC of 95.19 ± 0.08% on Real3D-AD and 94.03 ± 0.06% on the original 40-category Anomaly-ShapeNet benchmark, leading the reported point-level comparisons on both datasets. Object-level results rank first and second, respectively, among the compared methods. Module comparisons and mechanism analyses examine the contributions and limitations of reference routing and local-support correction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Point Cloud Anomaly Detection, Unsupervised Anomaly Detection, Local Feature Manifolds, Local Normal Support, Spatial-Context Gated Routing, Training-Free Methods

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_CV1JOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes BTM, a training-free memory-bank framework for 3D anomaly detection that addresses two limitations of conventional memory-based methods: mismatched normal references and the inability of nearest-neighbor distances to explicitly model allowable local geometric variation. BTM consists of three main components: Manifold-Aware Memory, which constructs local normal patterns from graph-geodesic neighborhoods of fixed FPFH features and stores local subspaces with finite tangential supports and orthogonal thicknesses; Spatial-Context-Gated Routing, which uses registered spatial information and multiscale context to select and rerank compatible references; and Corrected Tangent-Normal Scoring, which measures deviations beyond the support of the selected normal pattern.

The method requires only normal training point clouds and does not use learned encoders or synthetic anomalies. The authors report P-AUROC scores of 95.19±0.08% on Real3D-AD and 94.03±0.06% on the original 40-category Anomaly-ShapeNet benchmark.

Critical strengths

this paper proposes atraining-free framework using normal shae memory bank tas referencing for the anomal detection socring, ,The idea of using Manifold Aware Memory to build local normal patterns from grpah neueral ntwork is very interesting, and a novel idea comapred to ht enorma lpnomal point cloud input, mesh indeedd based on the geodesic feautre is moreaccurate theoritically.

  1. The paper use FPFH feature point learning featur e poitn descriptors, storing local feature subspaces,tangen and normal space of the mesh geometry topology feature, then ranks the most similar memory bank geometric features., This is vrey intersing no body has tried the poitn feature in this domain, to leverage the distinctive feature point-centered neighbouring point patch features, for memeory abnk establishemen.t
Critical weaknesses

I lik e the paper idea quite much, but only concern whether this draft version now stilll uncomplted draft version, as qualitative reuslts visually only presented in IFugre 2, only three object smapels, more qualitative reuslt shsodl be added , as I have not found the suppplemtary materials, no appendix pdf provided, and more viusal reuslts not provide in the paper.

Secondly, some technical details should be clarified more, as the curent veresion still ignore tehcnical detials, hard make ing reader to understand your implementations, In figure 2 top left, how do you conver t the g Mutual-kNN graph grah node feature to the normal bundles of frozen memory, what is the feaut redimension for the memory what is the tensor shape looking like, how is your Tangent-Normal proejction implemented? Like before/after projection feature dimension, an yoru model inpti can only be mesh what about poitn cloud input.

Some caption deitals are missing, hard to udnerstand the reustls, like what is the lien in fgure figur efor for (b) , better put anotaito nnext to (a) to (e lael, and why not show more latest baselien viusla reuslts like PO3AD, RD3D-AD baseliens, curently yo useem to cobine ablatio nviusla reuslts na dbaselie reuslts in one figure, too cpm[act, more visualabaltio nan dqualitative basleiene evluation reuslts shoudl be eprovided.

What is the runtime efficiency for various numbers of input points and the model complexity, like model parameter size? These mdoel model perofmrance deitals comapred baseliens and more implementation deitals shoudl be provide, and what is yiruo plan for th e public code realsing, I hva not doun any supplementary pdf reuslts, so may be oru papeexperiemnt sstill no t complted?

Questions for authors
  1. Could the authors clarify the precise methodological novelty of the proposed approach compared with existing 3D anomaly detection methods? In particular, what is the specific role and novelty of the proposed geometric metric encoding and topological consistency learning, beyond combining existing geometric features such as point positions, surface normals, and curvature?
  2. Could the authors clarify the completeness of the experimental evaluation? The current version provides limited qualitative results and does not appear to include supplementary materials or an appendix with additional experiments. Are there additional quantitative and qualitative results, particularly comparisons with recent strong baselines such as PO3AD and RD3D-AD, that are not included in the current submission?
  3. Could the authors provide more detailed ablation results to isolate the contribution of each geometric component? Specifically, how much do relative point positions, normal-vector variations, and curvature contribute individually and jointly to the final performance? This clarification is important for determining whether the reported improvements primarily come from the proposed method or from combining existing geometric features.
  4. Could the authors clarify the implementation details of the proposed feature construction and memory representation? In particular, how are the geometric Mutual-kNN graph features converted into the normal bundles of the frozen memory, what are the corresponding feature dimensions and tensor shapes, and how is the Tangent-Normal projection implemented? These details are important for assessing the technical contribution and reproducibility of the method.

I would be inclined to increase my score if the authors provide more convincing evaluation and detailed ablation and qualitative results during the rebuttal.

Reviewer_JLN4OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes BTM, a training-free memory-based framework for 3D anomaly detection using normal training point clouds and fixed FPFH features. It constructs local normal patterns using graph-geodesic neighborhoods and local PCA, selects references through spatial constraints and selective multiscale-context reranking, and scores deviations beyond estimated tangential support and orthogonal thickness. Experiments on Real3D-AD and the original 40-category Anomaly-ShapeNet benchmark report strong point-level results, together with module comparisons and mechanism analyses.

Critical strengths
  1. A coherent treatment of reference selection and normal variation. Separating compatible-reference retrieval from support-based scoring is practically motivated and gives an interpretable framework that does not require learned encoders or synthetic anomalies.

  2. Encouraging results within the evaluated setting. The paper reports strong point-level performance on two benchmarks and provides module comparisons and mechanism diagnostics. It also acknowledges limitations, including the modest contribution of context reranking and dependence on registration.

Critical weaknesses
  1. Limited methodological novelty. In my assessment, the contribution is largely incremental: the framework combines established geometric descriptors, graph-based neighborhoods, local PCA, spatially constrained retrieval, and support-based residual correction. The integration is reasonable, but the paper does not sufficiently establish a substantial methodological advance beyond these existing ingredients. In my assessment, the central contribution is closer to a practical heuristic refinement than a substantive methodological innovation.

  2. Incomplete reproducibility details. Although many core hyperparameters are provided, the category-specific gating thresholds used in the main experiments are only described as being stored in run configurations; their values and selection procedure are not specified in the paper. Registration and pose-selection settings and the visible-surface augmentation procedure also remain insufficiently detailed. These choices affect the reference memory and retrieval, making independent reproduction of the reported results difficult without complete configurations or an executable implementation.

  3. Limited evidence across feature representations. BTM is evaluated only with fixed FPFH descriptors. Comparisons against methods using other representations, such as the reported PatchCore–PointMAE baseline, do not substitute for controlled experiments that change the feature extractor within BTM. Evaluating alternative representations, for example PointMAE or PointBERT features, with appropriately matched controls would help establish whether the gains arise from the proposed memory and scoring design beyond this particular descriptor. The current results support the evaluated FPFH setting, but provide insufficient evidence for broader applicability. Together with the limited novelty, this leaves the contribution insufficient for acceptance in my assessment.

Questions for authors

Please refer to the Critical Weaknesses section.

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

The manuscript mainly focuses on training-free 3D point cloud anomaly detection using normal-only data, emphasizing local normal pattern modeling, spatial-context-aware reference routing, and support-based anomaly scoring for accurate anomaly detection and localization. It reports leading point-level results among the compared methods on Real3D-AD and Anomaly-ShapeNet, and examine module contributions through system comparisons, heldout representation analysis, and scoring diagnostics.

Critical strengths
  1. This manuscript proposes a training-free 3D anomaly detection framework that requires only normal point clouds, without learned encoders or synthetic anomalies, making the approach relatively simple and practical.

  2. The proposed method achieves leading reported P-AUROC results among the compared methods on public datasets: Real3D-AD and Anomaly-ShapeNet.

Critical weaknesses
  1. This manuscript lacks quantitative comparisons of inference time, memory consumption, and computational efficiency against competing methods. Given the claimed training-free advantage, reporting these metrics would help establish the practical efficiency and scalability of the proposed framework.

  2. The proposed method relies on a memory bank constructed from normal point clouds, but the manuscript does not establish how detection performance varies with the number of available normal reference samples. How many normal samples are required for reliable performance, and how does the method behave under more restrictive few-shot settings?

  3. The proposed method heavily relies on spatially constrained reference retrieval, making accurate registration critical. However, the manuscript neither directly evaluates registration quality nor comprehensively analyzes the impact of registration errors. A systematic sensitivity plot across a wider range of rotation and translation errors is needed to identify when performance degrades significantly.

  4. The proposed framework mainly integrates established techniques, including FPFH, neighborhoods, PCA, spatially constrained retrieval, and support-based scoring. The manuscript should more clearly distinguish its fundamental methodological innovations from the combination of existing components.

  5. Several related-work statements group multiple methods and place their citations at the end of each sentence, making the attribution of specific claims less explicit.

  6. The manuscript shows some tendency toward AI-assisted writing, as citations are consistently placed at the ends of sentences, even when multiple methods are discussed in a single sentence, resulting in several references being clustered together rather than directly linked to individual methods.

Questions for authors

I appreciate the motivation and the overall approach of building a normality model based on features extracted from normal samples. However, the major weaknesses outlined above remain my primary concerns, and I would appreciate clear explanations and responses to these issues.

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

Related papers

Discussion (0)

Sign in to comment.