acceptodds
Withdrawn from ICLR 2027

Learning Anomaly-Aware Normal Subspaces for Few-Shot Industrial Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Few-shot industrial anomaly detection aims to identify unseen defects us- ing only a few normal images. Recent reconstruction-based methods seek to model normality through feature restoration. However, they may fail to generalize to unseen normal variations or inadvertently reconstruct anoma- lous information, leading to false positives and false negatives.To address this issue, we propose NormProj, which learns an anomaly-aware normal subspace by determining which feature variations should be preserved or re- jected. An orthogonal-basis generator constructs the normal subspace from augmented support features, while joint normal reconstruction and paired synthetic-anomaly restoration encourage the preservation of legitimate nor- mal variations and the exclusion of anomaly-induced changes. After train- ing, the generator is discarded, enabling efficient projection-based anomaly scoring without an additional neural reconstruction network. Experiments on MVTec-AD and VisA demonstrate improved image-level anomaly de- tection over reconstruction- and PCA-based baselines, together with com- petitive localization performance across multiple few-shot settings. Code is available at https://github.com/sfhiswfgh/NormProj.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Few-Shot Anomaly Detection, Orthogonal Projection, Industrial Anomaly Detection, Visual Foundation Models, Representation Learning, Anomaly Localization

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

This work extends SubspaceAD [1] by introducing additional constraints on the separation between the normal subspace and anomalous features, where the anomalous features are generated using synthetic anomalies. The motivation is also closely related to PromptAD [2], which employs CLIP-derived semantic anomaly anchors to regularize the distance between normal and abnormal feature spaces, thereby improving few-shot industrial anomaly detection performance.

[1] SubspaceAD: Training-free few-shot anomaly detection via subspace modeling. CVPR 2026.

[2] PromptAD: Learning Prompts with only Normal Samples for Few-Shot Anomaly Detection. CVPR 2025.

Critical strengths

The paper is well organized and easy to follow. In addition, the method achieves competitive performance on industrial anomaly detection benchmarks by exploiting richer information conveyed by anomalous patterns.

Critical weaknesses

I discuss the weaknesses one by one below, in descending order of importance.

  1. The motivation is clear, however, the novelty of the proposed method is limited. The overall contribution is largely incremental, as the core idea has already been explored in prior studies, including SubspaceAD [1], GLAD [2], and PromptAD [3].

  2. The experimental evaluation is rather limited. More commonly used benchmarks, such as MPDD, Real-IAD, and datasets involving logical anomalies, should be included to further validate the effectiveness and generalizability of the proposed method.

  3. The paper has several formatting issues. For example, the font style and line spacing do not conform to the required formatting guidelines.

[1] SubspaceAD: Training-Free Few-Shot Anomaly Detection via Subspace Modeling. CVPR 2026.

[2] GLAD: Towards Better Reconstruction withGlobal and Local Adaptive Diffusion Models for Unsupervised Anomaly Detection. ECCV 2024.

[3] PromptAD: Learning Prompts with only Normal Samples for Few-Shot Anomaly Detection. CVPR 2025.

Questions for authors

See weaknesses.

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

The paper proposes NormProj for few-shot industrial anomaly detection. Given a frozen DINOv2-B encoder and a small set of normal support images, a category-specific MLP generates candidate bases that are averaged across augmented views and orthogonalized with QR. Training combines normal-feature reconstruction, paired synthetic-anomaly restoration, and a synthetic-mask localization loss. At inference, the generator is discarded; only the category mean and orthonormal basis are retained, and projection residuals are used for anomaly scoring. The paper evaluates on MVTec-AD and VisA at 1, 2, and 4-shot settings. It reports consistent image-level AUROC gains over its reconfigured SubspaceAD baseline, while localization results are mixed and the released code does not include the full training and evaluation pipeline.

Critical strengths
  1. The paper has a clear motivation and an intuitive geometric interpretation: the objective encourages the subspace to preserve normal variation while suppressing paired synthetic-anomaly changes. The projection-based inference representation is simple and compact, and the reported category state size is consistent with the stated dimensions.

  2. The reported image-level detection results are promising across both datasets and all three shot settings. In particular, the reported I-AUROC gains over SubspaceAD on VisA range from 2.20 to 2.63 percentage points. The paper also includes initial ablations of its losses, basis construction, and subspace rank.

Critical weaknesses
  1. Eqs. (1) and (3), without the localization loss in Eq. (4), imply a closed-form spectral baseline: the top-r eigenspace of 2C_u − C_delta. This baseline is missing from Table 3, and the Direct-Q setup is underspecified, leaving the generator’s independent contribution unclear.

  2. The anomaly-suppression mechanism is evaluated mainly with synthetic anomalies. Eq. (1) does not establish that these anomalies cover real defect directions or that unseen normal variations are preserved.

  3. Reproducibility is limited. Section 4.1 does not specify the validation data or metric for early stopping, and key augmentation and synthetic-anomaly details are missing. The released code excludes training and evaluation.

  4. Results are reported as three-seed means without uncertainty estimates or per-category results. NormProj also has lower P-AUROC than SubspaceAD in all six settings. In addition, the PCA gains stated in the discussion do not match Table 3.

Questions for authors
  1. Can you compare NormProj with the top-r eigenspace of 2C_u − C_delta, implied by Eqs. (1) and (3) without the Eq. (4) localization loss, and fully specify the Direct-Q baseline?

  2. What validation data and metric determine early stopping? Please provide the missing training and synthetic-anomaly details, full code, and per-seed/per-category results with uncertainty estimates.

  3. Can you clarify why the PCA gains reported in the discussion differ from the values computed from Table 3, and whether the PCA settings in Tables 1 and 3 are matched?

Reviewer_6mJaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes NormProj, a novel framework for few-shot industrial anomaly detection based on anomaly-aware normal subspace learning. Unlike traditional PCA-based methods, NormProj jointly preserves normal feature variations and suppresses anomalous changes through synthetic-anomaly supervision. Experiments on MVTec-AD and VisA demonstrate improved image-level anomaly detection performance with competitive anomaly localization results.

Critical strengths
  1. Clear motivation. The paper identifies meaningful limitations of existing reconstruction-based and PCA-based anomaly detection methods.
  2. Simple and effective design. The proposed framework combines normal reconstruction and anomaly exclusion in a relatively straightforward manner, supported by theoretical analysis.
  3. Promising experimental results. NormProj achieves strong image-level detection performance across different few-shot settings.
  4. Efficient inference design. The basis generator can be discarded after training, reducing the model's deployment footprint.
Critical weaknesses
  1. Limited methodological novelty. The proposed method mainly extends conventional subspace learning with synthetic-anomaly supervision. Its methodological contribution beyond existing approaches such as SubspaceAD could be further clarified.
  2. Dependence on synthetic anomalies. The method relies on artificially generated anomalies during training. It remains unclear how well the learned subspace generalizes to real anomalies that differ significantly from the synthetic ones.
  3. Inconsistent localization improvements. Although image-level detection performance improves, pixel-level AUROC is slightly lower than SubspaceAD across the reported settings, suggesting limitations in localization accuracy.
  4. Insufficient efficiency evaluation. Although the paper emphasizes efficient inference, a direct comparison of inference latency and total computational costs against competing methods would strengthen this claim.
Questions for authors
  1. Could you clarify more concretely what is fundamentally new in NormProj compared to existing subspace-based methods such as SubspaceAD?
  2. Since the framework relies on synthetic anomalies for anomaly-aware subspace learning, how sensitive is NormProj to the choice and distribution of these synthetic anomalies? Have you tested scenarios where the real anomalies at test time differ substantially from the synthetic ones, and can you provide additional experiments or analysis on such distribution shifts?
  3. Can you provide more insight into why localization does not benefit as much from NormProj, and whether any variants could improve localization?
  4. Could you include a more detailed efficiency comparison against baselines to substantiate the efficiency claims?

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

Related papers

Discussion (0)

Sign in to comment.