acceptodds
Withdrawn from ICLR 2027

UniNorm: Normal-Direction Surface-Deviation Measurement for Multi-Category 3D Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-category 3D anomaly detection must detect and localize defects across heterogeneous categories using only normal training data. Existing methods largely improve representations of normality while inheriting feature distance, reconstruction error, or field residual as anomaly evidence, coupling representation quality with the suitability of the resulting measurement. Under matched registered inputs and identical object pooling, we compare geometric residuals with learned implicit-field residuals and find performance differences. We therefore introduce UniNorm, which formulates template-aligned 3D anomaly detection as normal-direction surface-deviation measurement. UniNorm uses absolute normal projection to retain surface-normal departure while suppressing tangential correspondence motion, and optionally applies a nonparametric normal-dispersion correction (NULB) together with a shared contrast-tail object readout. The scorer has no learned parameters; only a lightweight category-specific nominal template is stored. On Anomaly-ShapeNet and Real3D-AD, UniNorm achieves 92.60/90.53% and 85.14/83.48% object-/point-level AUROC, respectively. With registration and object pooling matched, absolute normal projection improves over a class-conditioned implicit-field residual by 6.83/8.95 and 3.16/8.26 O/PAUROC points on the two benchmarks. These results highlight anomaly measurement as a first-class design choice in template aligned 3D anomaly detection.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
3D anomaly detection, industrial inspection, point cloud, surface-deviation measurement

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_QKJhOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes UniNorm, a lightweight multi-category 3D anomaly detection framework that quantifies anomalies as surface-normal deviations relative to category-specific templates. It aligns test point clouds via RANSAC-ICP, scores samples via absolute normal projection, and aggregates results with a shared contrast-tail readout — the entire scoring pipeline is parameter-free. Built on the core idea of decoupling anomaly measurement from normality representation, UniNorm outperforms SDF-based baselines on Anomaly-ShapeNet and Real3D-AD under matched settings. Ablation and robustness analyses further validate its efficacy and clarify its limitations.

Critical strengths

This work draws a valuable conceptual distinction between the representation of normal geometry and the measurement of anomalies. Rather than introducing additional complex learned representations, UniNorm quantifies geometric deviations by projecting point displacements onto surface normals. This approach is conceptually simple and intuitive, with a clear geometric interpretation. Another key strength of the paper is its rigorous controlled experimental design. The authors hold registered inputs and object-level aggregation constant when comparing different residual formulations, ensuring fair and consistent comparison conditions. Component ablation studies further disentangle the contributions of normal and tangential displacements, making the findings more convincing than those derived from standard benchmark comparisons alone.

Critical weaknesses

(1) The core experimental claim remains incompletely substantiated. Although the authors control for registration and object-level aggregation, the underlying representational paradigms of UniNorm and the SDF baselines remain fundamentally distinct. It is therefore difficult to ascribe the observed performance gains exclusively to the proposed anomaly measurement approach, as discrepancies in representation and training may also contribute to the performance gap. Additional controlled comparisons are required to more rigorously isolate the specific effect of the measurement metric itself.

(2) The methodological novelty of the work appears relatively modest. The core idea of using surface-normal deviation bears close resemblance to classical point-to-plane distance metrics, and as the authors themselves acknowledge, this does not constitute a fundamentally new geometric primitive. The primary contribution of the work is thus largely empirical rather than methodological in nature.

(3) The proposed NULB correction yields only marginal improvements over absolute normal projection, and fails to consistently enhance localization performance across all defect categories. Notably, its performance degrades substantially for defects featuring broken or thin boundaries, which casts doubt on its overall practical utility.

(4) The method’s strong reliance on accurate registration and sufficient template coverage limits its robustness in practical deployment scenarios. Furthermore, the NULB coefficient is selected using labeled Anomaly-ShapeNet test data rather than a held-out validation set, which raises concerns regarding potential test-set tuning and the reliability of the reported generalization results.

Questions for authors

Clarification Requests: (1) The NULB coefficient (β\beta) was selected using labeled Anomaly-ShapeNet test results. Could the authors clarify how much the main Anomaly-ShapeNet conclusions depend on this test-set tuning, e.g., by reporting the corresponding results with β\beta=0 or with a parameter choice made independently of the test labels? (2) The main residual comparison controls registration and object-level pooling, but the SDF-based methods and UniNorm still use different underlying representations. Could the authors clarify the intended scope of the conclusion that normal-direction measurement is superior—i.e., whether the claim is limited to the evaluated SDF probes or is meant to be more general? (3) Could the authors clarify the frequency and impact of registration failures in the standard evaluation, rather than only in the stress tests? In particular, are all failed or low-quality registrations included in the reported headline results for both datasets?

Suggestions for Future Work: It would be useful to investigate extensions that are less dependent on accurate registration and that can handle fully missing surface regions, for example through more robust alignment or bidirectional/surface-completeness-aware anomaly measures.

Reviewer_mf7zOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a simple 3d-registration-based method to conduct 3d anomaly detection. The idea is to compare the test 3d point cloud and a normal 3d object template and calculate the surface deviation. The proposed method achieve a competitive performance on standard 3d anomaly detection datasets as other methods while being efficient since it is parameter-efficient.

Critical strengths

The proposed method is easy to understand and direct, which should be easy to implement.

The method shows good object-level performance on standard 3d anomaly detection and localization benchmarks with real-time inference capability.

Critical weaknesses

The paper presentation is bad. The overview figure looks to be AI generated which includes the illustration of the object template and the test object point cloud. I would recommend the authors to remake them.

Also, the motivation illustrated in figure 1 is not clear to me. I am confused on the inconsistency of the annotated defect regions in the three subfigures.

For the proposed method, it looks to be classical 3D registration, which offers no design related to "representation learning" for the conference topic.

Some problem formulation parts are not clear. For example, from my understanding this work targets anomaly detection of 3d point clouds under multi-category setup yet the method requires object-specific alignment.

Questions for authors

Please check the weaknesses section.

I am confused on the "multi-category setup" of the task while the model requires object template for matching. Please clarify on this point.

Reviewer_J5UHOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a nonparametric method for multi-category 3D anomaly detection using normal-direction deviations from category-specific templates. The method aligns a test point cloud to a category-specific normal template using RANSAC–ICP and computes absolute normal projections of nearest-neighbor displacements. An optional normal-dispersion correction, NULB, suppresses potential tangential leakage, while a contrast-tail statistic aggregates point scores into an object score. The scoring rule and hyperparameters are shared across categories, with no learned scoring parameters. Experiments on Anomaly-ShapeNet and Real3D-AD show competitive results.

Critical strengths

The method combines a simple geometric scorer with category-specific templates and achieves competitive detection performance.

Critical weaknesses
  1. The authors selected the NULB coefficient β\beta using labeled Anomaly-ShapeNet test results rather than a separate validation set (Line 772-773).
  2. Several figures, particularly Figures 2–4, appear overly compressed, with text and annotations that are difficult to read at the intended viewing size. This makes the pipeline harder to follow and the plotted results and qualitative comparisons harder to assess.
Questions for authors

Given that normal projection is well established and NULB offers only modest gains while hurting localization for some defect types, could the authors clarify what constitutes the main methodological novelty beyond an effective geometric baseline?

Could the authors clarify whether generative AI tools were used to create Figure 2 and, if so, describe how they were used?

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

Related papers

Discussion (0)

Sign in to comment.