acceptodds
Desk rejected at ICLR 2027

PointSimAD: Training-free Zero-shot 3D Anomaly Detection via Intrinsic Self-similarity

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Zero-shot 3D anomaly detection (ZS-3DAD) aims to detect and localize anomalies in unseen object categories without requiring target-category training samples, making it particularly attractive for data-scarce industrial inspection. However, existing ZS-3DAD methods often rely on external auxiliary data to adapt pretrained models to the industrial domain, making their performance sensitive to the distribution of such data. In this paper, we propose PointSimAD, a training-free framework for ZS-3DAD that requires no fine-tuning with external industrial data and directly exploits the intrinsic structure of target point clouds. Specifically, we introduce a structural fingerprint that encodes local geometry, global shape, and spatial distribution into a sample-level descriptor. Based on this structural fingerprint, a structure-aware point grouping module adaptively determines the local structural granularity and partitions the point cloud into local point sets. A frozen 3D encoder then encodes the resulting local point sets into patch-level feature representations. By comparing each patch feature with similar features within the same object, we assess intra-object structural consistency and detect and localize anomalies based on the resulting local inconsistency. Experiments on three 3D anomaly detection benchmarks demonstrate that PointSimAD achieves strong and consistent performance, particularly for point-level anomaly localization, while requiring substantially lower memory and computation than existing adaptation-based approaches. The source code will be made publicly available upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
zero-shot 3D anomaly detection, training-free

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ER2hOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PointSimAD detects anomalies in a single point cloud without target-category normal references or task-specific training. It uses handcrafted geometric statistics to select one of three grouping configurations, extracts patch features with a frozen ULIP-2 PointBERT encoder, and subtracts the feature-wise median across patches. Each patch is scored by its average cosine distance to its four nearest feature neighbors within the same object. Scores are interpolated to the original points for localization and aggregated for object-level detection. The paper evaluates the method on Real3D-AD, Anomaly-ShapeNet, and MulSen-AD.

Critical strengths
  1. The training-free, reference-free setting is practically useful. The method avoids collecting normal examples from each target category and uses a relatively simple inference pipeline. Its reported memory footprint and throughput are encouraging, subject to the timing scope discussed below.

  2. Point-level localization is promising. Table 1 reports P-AUROC values of 86.4, 88.5, and 81.5 on the three datasets, respectively, the highest among the compared methods. Figure 4 shows the resulting anomaly maps, and Table 3 provides quantitative evidence that median centering and adaptive grouping improve the reported baseline. These results give the approach practical value despite the concerns below.

Critical weaknesses
  1. The adaptive grouping mechanism is not sufficiently justified, which limits the methodological contribution. The method combines geometric summary statistics, a frozen encoder, median centering, and nearest-neighbor scoring. Its adaptive component makes a hard minimum-cost choice among three presets. The paper does not establish why matching the proposed fingerprint should identify a configuration that is better for anomaly detection. Bounding-box ratios, PCA eigenvalue ratios, and radial dispersion describe global properties; their additional value for selecting a configuration suited to local surface variation remains unclear. My concern is their contribution beyond local distance statistics. Whole-module ablations in Table 3 do not answer this question, and the robustness of the fixed per-dimension weighting is not examined.

    The sample-level evidence is also incomplete. Fixed-Fine is worse than both Fixed-Medium and Fixed-Coarse on all four metrics in Table 5. Figure 5 shows three examples but does not give their consistency costs or the actual SAPGM choices. Selection counts and proportions are missing. The configurations change the number of centers while keeping 64 points per patch, four feature neighbors for scoring, and eight spatial neighbors for interpolation. Fine's weaker performance could partly reflect the resulting changes in reference redundancy and interpolation support. The current experiments do not separate this interaction with the scorer from the choice of geometric scale. Before/after SCDM diagnostics or examples of the actual reference patches would also help establish the proposed explanation beyond its aggregate ablation gain.

  2. The implementation and cost reporting leave the practical contribution difficult to assess and reproduce. Section 3.2 and Appendix A require FPS center selection and kNN grouping for each candidate, so all three configurations need candidate statistics before selection. This does not remove the initial FPS downsampling to 16,384 points. Algorithm 1 lists another FPS/grouping step after selection without explaining whether candidate results are reused. Caching or shared sampling could reduce this cost, but the implementation is not described. Table 2 and Figure 3 exclude preprocessing and do not isolate candidate preparation or clearly establish whether it is timed. Table 5 has no per-configuration latency or memory measurements.

    This omission matters because adaptive grouping improves O-AUROC over Fixed-Coarse by only 0.6 and 0.7 percentage points; P-AUROC gains over the best fixed configuration are 0.6 and 2.2 points. The broader advantage is mainly in localization: object-level AUROC is 79.4 versus 84.1 for GS-CLIP on Anomaly-ShapeNet, and 72.4 versus 74.5 for BTP on MulSen-AD. These results need to be assessed alongside the cost of selection.

    The most consequential reproducibility gap is the candidate fingerprint in Eq. (3). Table 7 lists the statistics, but “center-set and local-group statistics” does not define, dimension by dimension, which points are used or how measurements across groups are aggregated. Section 4.2 says five dimensions, whereas Section 3.2 and Appendix A say twelve. Other missing details include the coordinate normalization and 4,096-point subsampling rule, exact checkpoint and L2 feature extraction interface, Gaussian interpolation bandwidth and weights, any score normalization, and the within-category construction of P-AUROC before the stated category macro-average. The grouping used when SAPGM is disabled in Table 3 is also unidentified; its SCDM-only result of 75.0/83.8 on Real3D-AD matches none of the fixed configurations in Table 5.

I lean toward rejection in the current form. The localization results are useful, but the link between the selection criterion, the configuration chosen for an individual sample, and the resulting detection quality and cost is not established. Resolving these points would affect my assessment more than further aggregate benchmark results.

Questions for authors
  1. Clarification of the selection rule: Please define every component of the candidate fingerprint in Eq. (3), including the point set, normalization, and aggregation across local groups. Which fingerprint dimension was used in the experiments, five or twelve?

  2. Evidence for the fingerprint design: What do the shape and spatial-distribution features add beyond local statistics? A controlled comparison of local-only, local-plus-shape, local-plus-distribution, and the full fingerprint would directly address my main concern. Please report both performance and changes in the selected configurations, including sensitivity to the weighting choice.

  3. Evidence for the configuration choices: For representative samples where Fine, Medium, and Coarse respectively work best, can you provide the three consistency costs, the actual SAPGM choice, and the localization results? How often is each configuration selected? Can the poorer Fine results be explained by sensitivity to the four-neighbor scorer or eight-center interpolation? A few controlled diagnostics, including SCDM before/after results or selected reference patches, would help distinguish these explanations.

  4. Clarification and measurement of cost: Does the timing include initial downsampling, fingerprint extraction, and all three candidate FPS/kNN operations? Are candidate centers and groups cached or shared, and is the selected configuration recomputed as Algorithm 1 appears to indicate? Accuracy, latency, and peak memory for the three fixed configurations and SAPGM under the same conditions would establish whether selection provides a useful trade-off.

  5. Clarification of reproducibility details: Please provide the exact checkpoint and feature extraction settings, coordinate normalization, fingerprint subsampling rule, Gaussian interpolation parameters, score normalization, and point-level evaluation procedure. How are point scores and labels combined within a category before computing P-AUROC?

  6. Clarification of the ablation baseline: What fixed grouping and other settings are used for the baseline and SCDM-only rows of Table 3? Please explain their relationship to the fixed-configuration experiments in Table 5.

Reviewer_1TNbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a training-free method for zero-shot 3D anomaly detection. It constructs a geometric structural fingerprint to select a point-grouping configuration, then extracts patch features using a frozen ULIP encoder. The method detects anomalies by comparing each patch with similar patches in the same object, then uses these scores to locate defects and identify anomalous objects. Experiments on three datasets show good results.

Critical strengths

The proposed method is memory efficient compared to existing works.

Critical weaknesses

The authors first claim adaptive grouping as a way to address differences in local receptive fields (L154–160), but in practice they only vary the number of patch centers while keeping each neighborhood fixed at 64 points, which seems to mainly change patch coverage and overlap (L300–303). It is unclear how this configuration addresses the stated receptive-field issue, rather than primarily changing patch coverage and overlap. In Table 3, the SCDM-only variant appears to use fixed grouping, yet its results do not match any of the fixed-grouping configurations in Table 5. The paper does not explain this difference, making it difficult to determine whether the gains come from adaptive grouping or other changes in the experimental settings.

Questions for authors

Have the authors tried different neighborhood sizes beyond 64 points? Varying the neighborhood size would more directly test the role of local receptive fields and help distinguish its effect from that of changing the number of patch centers.

Which grouping configuration is used for the SCDM-only variant in Table 3, and why do its results differ from all fixed-grouping results in Table 5? Are the other experimental settings identical?

Reviewer_WN6NOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PointSimAD, a training-free zero-shot 3D anomaly detection framework that requires no normal reference samples of the target category. Unlike existing methods that rely on external industrial data to adapt pretrained models, PointSimAD directly leverages the structural redundancy within a single test object to establish a reference for anomaly detection. The method mainly consists of three parts. First, Multi-level Structural Fingerprint extracts geometric statistical information from three aspects—local scale, global shape, and spatial distribution—to guide point cloud partitioning. Second, SAPGM adaptively selects the local structural granularity suitable for the current object among three predefined configurations: coarse, medium, and fine. Third, intrinsic self-similarity modeling: frozen ULIP-2 PointBERT is used to extract Patch features, common features are removed through Structural Consensus Decomposition (SCDM), and then the cosine distance between each Patch and the four most similar Patches within the object is calculated to obtain anomaly scores. The authors evaluated the method on three benchmark datasets, Real3D-AD, Anomaly-ShapeNet, and MulSen-AD. PointSimAD achieved the highest point-level AUROC among the listed methods on all three datasets, namely 86.4%, 88.5%, and 81.5%, respectively. The paper also reports good inference efficiency: on Real3D-AD, the peak GPU memory usage is 498 MB, and the throughput is 14.2 FPS.

Critical strengths
  1. The zero-shot anomaly detection method in this paper is training-free and does not rely on additional auxiliary data. The proposed framework leverages the structural self-similarity within an object and constructs an intrinsic reference solely from a single test point cloud. The problem studied is of certain value.
  2. The method achieves good performance, obtaining promising detection results on the three evaluation benchmark datasets. In addition, it also has good inference efficiency.
Critical weaknesses
  1. Experimental validation is insufficient.

  2. The paper presents inconsistencies in the specific settings of the method. For example, Section 4.2 of the main text describes the structural fingerprint as five-dimensional, whereas Section 3.2, Equations (1)–(2), and Appendix A all define it as twelve-dimensional.

  3. The method proposed in the paper can essentially be regarded as computing anomaly scores through the four most similar Patches within the same object. However, normal structures are not necessarily repetitive. For example, for irregular or asymmetric objects, some completely normal local structures may appear only once, making it impossible to find similar internal references and thus causing false detections. For instance, on the Real3D-AD dataset, the O-AUROC for the seahorse category is only 67.5%, and the O-AUROC for the starfish category is only 66.1%, slightly above random guessing.

Questions for authors

Q1. Robustness of the intra-object self-similarity assumption. The proposed method assumes that normal local structures can find similar counterparts within the same object. However, this assumption may not hold for objects with unique or asymmetric normal structures. Furthermore, repetitive anomalous regions may serve as references for one another, potentially leading to underestimated anomaly scores. Could the authors discuss how the method handles these cases and provide further analysis or experiments? In particular, could the authors investigate the relatively poor performance on the Seahorse and Starfish categories of Real3D-AD and clarify whether it is related to the limitations of intra-object self-similarity? Q2. Inconsistency in structural fingerprint dimensionality. Section 4.2 states that the structure-aware router uses a five-dimensional geometric fingerprint, whereas Section 3.2, Eqs. (1)–(2), and Appendix A define a 12-dimensional fingerprint. Could the authors clarify which dimensionality was actually used in the reported experiments and explain this inconsistency?

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

Related papers

Discussion (0)

Sign in to comment.