acceptodds
Under review as a conference paper at ICLR 2027

Capacity-Constrained Global Normality Association with Transport-Conditioned Residual Calibration for Few-Shot Industrial Anomaly Detection

Abstract

Few-shot industrial anomaly detection (FS-IAD) identifies defects from a few normal reference images. Despite advances in pretrained representations, locally plausible matches need not form a jointly consistent normal explanation, while residuals conflate defects with association ambiguity and normal appearance drift. We propose Global Optimal-Transport Memory Calibration for Anomaly Detection (GOMC-AD), a training-free framework for capacity-constrained normality explanation. Optimal-Transport Guided Global Memory Association (OT-GMA) couples query patches through prescribed normal capacities and a fixed rejection budget, making reference allocation part of the anomaly criterion. Transport-Residual-Aware Test-Time Gaussian Calibration (TR-TGC) incorporates transport-derived association variance into heteroscedastic Gaussian scoring and removes its modeled contribution from historical second-moment updates. Asymmetric feedback reallocates rejection mass while retaining the first-pass normal moments used to define residuals. Across MVTec AD, VisA, and Real-IAD with 1-, 2-, and 4-shot support, GOMC-AD outperforms the compared methods in mean image- and pixel-level area under the receiver operating characteristic curve (AUROC). Image-level AUROC improves by up to 5.3 percentage points and pixel-level AUROC by up to 1.3 points over the strongest compared methods. An anonymous project page is available at https://anonymous.4open.science/r/GOMC-AD-C25C/.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
few-shot anomaly detection, industrial anomaly detection, optimal transport, test-time adaptation, uncertainty calibration

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3wSkOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GOMC-AD for few-shot industrial anomaly detection, comprising Optimal-Transport Guided Global Memory Association (OT-GMA) and Transport-Residual-Aware Test-Time Gaussian Calibration (TR-TGC). OT-GMA jointly matches query and support patches through normal-capacity constraints and a fixed rejection budget. TR-TGC uses association variance and historical residual statistics for scoring, together with feedback and online updates. Experiments on MVTec AD, VisA, and Real-IAD under 1-, 2-, and 4-shot settings report higher image- and pixel-level AUROC than the compared methods, with ablation studies examining the contributions of the components.

Critical strengths
  1. Joint matching constrains the reuse of normal references. Shared capacities couple assignments across query regions, providing a mechanism for detecting globally inconsistent configurations even when individual patches have plausible normal matches.
  2. Residual calibration accounts for both association uncertainty and normal appearance variation. It combines transport-derived variance with historical residual statistics and preserves the normal reference during feedback. Targeted ablations and results across three datasets support these design choices.
Critical weaknesses
  1. The specific contribution of shared capacities is not fully isolated. Existing ablations establish the effectiveness of OT-GMA as a whole, but do not separately quantify the benefit of its central capacity-competition mechanism.
  2. Quantitative comparisons with closely related methods remain limited. In particular, FOCT also combines transport-based matching with online adaptation. Although the paper discusses their methodological differences, the lack of a direct experimental comparison limits the assessment of the gains over this family of methods.
Questions for authors
  1. Can the authors provide a targeted ablation showing how much improvement comes specifically from the shared-capacity constraints?
  2. Can the authors provide a quantitative comparison with FOCT under compatible settings and include additional representative matching or adaptation methods?
  3. Are rotation and flip augmentations applied to all categories? For categories with a consistent orientation or spatial layout, how does disabling these augmentations affect detection performance?
Reviewer_6yBCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GOMC-AD for few-shot industrial anomaly detection. It combines capacity-constrained optimal transport matching with online Gaussian residual calibration using frozen DINOv3 features. A second transport step further refines rejection assignments while retaining the original normal references. Experiments on three datasets under 1-, 2-, and 4-shot settings show that the full online model achieves the highest reported average AUROC, with improvements of up to 5.3 and 1.3 percentage points in image-level and pixel-level AUROC, respectively.

Critical strengths

1.The shared capacity constraint helps prevent multiple patches from repeatedly matching the same normal reference. Keeping the original normal references unchanged during online updates also provides a stable basis for residual scoring.

2.The experiments cover both anomaly detection and localization across several benchmarks and few-shot settings. The component ablations and the comparison between static and streaming settings help clarify the effects of the proposed design and online adaptation.

Critical weaknesses
  1. The current ablations do not clearly show which parts of the method are responsible for the improvements. OT-GMA changes view selection, spatial guidance, soft matching, and support capacity together, making it difficult to assess the contribution of the capacity constraint alone. The subtraction of association variance from historical statistics is also not tested separately. More targeted ablations would help establish the value of these two components.

  2. I have some concerns about the fairness of the comparison. GOMC-AD uses information from previously processed unlabeled test images to update its statistics, whereas most of the baselines do not. It is therefore hard to tell how much of the reported improvement comes from the proposed matching method and how much comes from online adaptation. A comparison under the same online setting would make the results more convincing.

Questions for authors
  1. FOCT is closely related to this work, but despite the discussion of their differences, Table 1 does not include a direct comparison. I would like to see both methods evaluated with the same features, resolution, and support sets. It would also help to separate the effects of fixed reference capacities, the rejection column, and view-specific transport in the ablation study.

  2. I am concerned about whether the main comparison is entirely fair. GOMC-AD updates its statistics using previously processed test images, but it is unclear whether the baselines have access to the same information. This matters because Table S8 shows a noticeable improvement from online adaptation, with image AUROC increasing from 90.3 to 95.2 on VisA in the 1-shot setting. The authors should clarify how much of the gain comes from the matching method itself. I would also be interested in the effect of test image order, especially when several anomalous images appear consecutively.

  3. The capacity constraint seems useful for detecting missing or repeated components, but I wonder how it handles normal changes in object pose or component arrangement. Could these variations lead to false alarms? Some results on normal variations and defects of different sizes would help clarify whether the fixed rejection budget works reliably across different cases.

  4. Matching uncertainty is used for both anomaly scoring and updating historical statistics, but the current experiments do not separate these two effects. I would like to see an ablation of each use. It is also worth checking whether matches with higher uncertainty actually show greater variation in normal residuals, as this would help justify the uncertainty modeling.

Reviewer_8EZMOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes GOMC-AD for training-free few-shot IAD with frozen DINOv3 features. OT-GMA matches query and support patches by per-view entropic OT with fixed normal capacities and a fixed rejection budget. It uses the cost gap over nearest-neighbor matching, the rejection allocation, and the transport barycenter and variance as anomaly evidence. TR-TGC scores barycentric residuals with a heteroscedastic Gaussian and updates its statistics online from previously scored test images. The method reports the best mean AUROC in all 18 settings on MVTec AD, VisA and Real-IAD.

Critical strengths
  • Independent retrieval allows unrestricted reuse of support patches, so missing or duplicated parts go unpenalized.
  • One transport plan yields the discrepancy, the reference and its uncertainty, and scoring and history updates share one variance decomposition.
  • Transparent appendix (algorithms, constants, static vs. streaming results). Even the static variant leads on Real-IAD.
Critical weaknesses
  • Creating competition through fixed column quotas in balanced entropic OT, with transport-weighted pooling as the patch-level readout, already appears in FGPT (Wang et al., IEEE TAI 2026), which calls it a "competitive allocation mechanism" for patch-level AD. The dustbin follows SuperGlue. FOCT already scores anomalies by a transport-weighted matching cost (Table S7), and FastRef (CVPR 2026) also uses OT for few-shot IAD. Please state what is conceptually new beyond FGPT's competition mechanism and these precedents, and isolate it.

  • The "Baseline" in Table 3 (91.4 on MVTec AD 1-shot) is undefined OT-GMA never beats the strongest baseline by more than 0.1. Without streaming, adding TR-TGC lowers I-AUROC by 1.1–2.0 on MVTec AD/VisA.

  • Main results update statistics from previously scored test images, while all baselines score each query independently. Without this, GOMC-AD is below the strongest reproduced baseline in all 12 MVTec AD/VisA settings (Table S8; e.g., 94.6 vs. 97.0 I-AUROC on MVTec AD 1-shot). The "all 18 settings" claim therefore depends on the test stream. This is essentially few-shot online AD (FOADS; Wei et al., 2024). Please label the protocol, add the static row to Table 1, and compare under matched access (FOCT, MuSc, or baselines given the same online update).

  • Only AUROC is reported, and pixel AUROC is saturated (gains of 0.3–0.7 on MVTec AD/VisA), so PRO, AP and F1-max are needed. Reproduced baselines fall below published numbers (VisA 1-shot: AnomalyDINO 82.1 vs. 87.4; SubspaceAD 91.9 vs. 93.3). Several hyperparameters (α, λ's, ℓ, u, τ, pooling) and the "independent validation data" are unspecified. The sensitivity results also show unexplained cliffs, e.g., full-moment feedback costs 6.6 points at η = 0.03.

Questions for authors
  • FGPT (IEEE TAI 2026) already creates competition through fixed column quotas in balanced entropic OT and obtains patch-specific prompts by transport-weighted pooling. What is conceptually new beyond this? Starting from an FGPT-style barycenter readout, how much do the cost gap, the rejection column and the Gaussian each add?

  • Eq. (2) is balanced, so every support patch must be fully used, and even a normal query must account for all content of one support view. Why is full use needed rather than a cap? Please report results with a support-side dustbin or unbalanced OT.

  • Without test-stream updates, GOMC-AD is below the strongest reproduced baseline in all 12 MVTec AD/VisA settings (Table S8). Does it still lead when all methods have the same access to test data?

Comments on the paper

1 comment
Area_Chair_y57mOpenReview ↗

A Gentle Reminder: About One Day Left in the Discussion Period

Dear Reviewers,

This is a gentle reminder that approximately one day remains in the discussion period. Please review the authors' responses and other reviewers' comments, and engage in the discussion to clarify which concerns have been addressed and which remain unresolved.

Thank you for your time and effort.

Best,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.