acceptodds
Withdrawn from ICLR 2027

Prototype–GVA–IDAC: Prototype-Based Localization and Image-Level Discrimination for Cross-Dataset Industrial Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Industrial visual anomaly detection models are required not only to localize subtle defects in unseen products and acquisition environments, but also to provide reliable image-level predictions. However, appearance shifts in cross-dataset settings can weaken the correspondence between textual semantics and local visual evidence. Moreover, MultiADS directly averages the global text-based anomaly probability and the maximum response of the anomaly map, making it difficult to simultaneously achieve fine-grained localization and stable discrimination. To address these limitations, we propose Prototype–GVA–IDAC, built upon MultiADS with frozen CLIP encoders. Prototype enhances multi-level patch representations through soft prototype assignment, contextual reconstruction, and residual refinement, enabling local anomaly cues to transfer more robustly to unseen datasets. GVA independently aggregates global image features and multi-level local statistics, thereby providing complementary visual discriminative evidence for image-level prediction. IDAC separately calibrates normal and anomalous text anchors under identity-preservation and normal–anomaly separation constraints, reducing the discrepancy between fixed textual semantics and target-domain images. The model is trained on MVTec AD and evaluated on five datasets—BTAD, MPDD, KSDD2, VisA, and WFDD—without using any target-domain data for training intervention, hyperparameter selection, or threshold tuning throughout the entire process. Experimental results demonstrate that, under the same evaluation protocol, the proposed model exhibits clear advantages in cross-dataset defect localization and anomaly discrimination.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
industrial anomaly detection, zero-shot anomaly detection, cross-dataset generalization, pixel-level anomaly localization, vision-language models, prototype learning, CLIP adaptation

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_953kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Prototype–GVA–IDAC extends frozen-CLIP MultiADS with image-conditioned prototype refinement, a separate visual image-scoring branch, and bounded text-anchor residuals. Training uses MVTec AD supervision, and evaluation covers five target datasets. The paper reports three-seed structural ablations, scoring controls without test-batch statistics, and inference costs.

Critical strengths
  1. The complete eight-configuration ablation and fixed-scale scoring control strengthen the validation. They support complementary localization and image-scoring benefits and show that the main image-level gain is not simply caused by test-batch normalization.
  2. The same-batch, three-seed comparison gives consistent improvements over reproduced MultiADS across the five evaluated targets, with a measured and relatively modest inference-cost increase.
Critical weaknesses
  1. The targets are not development-independent: Section 4.1 and Section 5 explicitly acknowledge that historical target feedback informed structural choices. This conflicts with the abstract's stronger exclusion of target-domain intervention and hyperparameter selection. Fixed reruns and additional seeds cannot establish unbiased transfer to previously unseen development targets, which is central to the cross-dataset claim.
  2. The main new modules integrate established contextual aggregation and separate scoring mechanisms, but the paper does not isolate Prototype against matched-capacity alternative context blocks. The closest AA-CLIP and AdaptCLIP comparisons are also untested. The evidence for a distinctive methodological advance is therefore weaker than the evidence for improvement over MultiADS.
Questions for authors
  1. Which decisions used target feedback, and can the finalized configuration be evaluated on a target excluded from all development decisions?
  2. Does Prototype retain its advantage against a matched-capacity alternative context block, and how does the full method compare with the closest adaptation baselines under the same supervision protocol?
Reviewer_GqwpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Prototype-GVA-IDAC, a method for cross-dataset industrial anomaly detection. It combines prototype-based feature refinement, global-local visual scoring, and text-anchor calibration. Experiments on five target datasets show improvements over the MultiADS baseline in both anomaly localization and image-level detection.

Critical strengths
  1. Good Performance. The proposed method achieves relatively strong performance across five target datasets.
  2. Effective Prototype Refinement. The Prototype module demonstrates notable improvements, suggesting that contextual feature aggregation can be beneficial for this task.
  3. Easy to Follow.
Critical weaknesses
  1. Limited Novelty. The proposed method mainly combines existing techniques with incremental modifications. While Prototype provides potentially useful contextual feature refinement, it is conceptually related to INP-Former [1]. GVA essentially introduces a conventional visual classification branch, while IDAC performs lightweight text feature adaptation, closely related to AA-CLIP [2]. Overall, the framework offers limited methodological novelty.

  2. Lack of a Unified Methodological Motivation. The proposed modules address largely independent aspects of anomaly detection, but the paper does not clearly explain why these particular components should be combined or how they complement each other. Although the ablation results demonstrate performance improvements, the rationale behind their combination and the potential benefits beyond their individual contributions remain unclear.

  3. Limited Cross-Dataset Evaluation. Unlike MultiADS that evaluates multiple source-target dataset combinations, this paper only uses MVTec AD as the source training dataset. Additional experiments using other source datasets, such as VisA or MPDD, are needed to better demonstrate the cross-dataset generalization capability of the proposed method.

  4. Lack of Cross-Framework Validation. All the experiments are conducted exclusively on MultiADS, making it unclear whether the proposed modules are generally effective or specifically tailored to this baseline. Evaluating the key components on another related framework, such as AA-CLIP [2], would provide stronger evidence of their broader applicability.

  5. Insufficient Analysis of the Prototype Mechanism. While Prototype contributes most of the performance improvements, the paper lacks sufficient analysis of its mechanism. Comparisons with some adapters or alternative attention-based modules, along with ablations on the gated residual refinement, would help clarify whether the improvements stem from the proposed prototype mechanism rather than simply increased model capacity.

[1] Wei Luo, et al. Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection, CVPR, 2025.

[2] Wenxin Ma, et al. AA-CLIP: Enhancing Zero-Shot Anomaly Detection via Anomaly-Aware CLIP, CVPR, 2025.

Questions for authors
  1. Prototype Mechanism. What is the key difference between the proposed Prototype module and existing within-image prototype methods such as INP-Former [1]? What evidence supports the necessity of the specific prototype aggregation and gated refinement design, rather than a conventional attention or residual adapter?

  2. Module Complementarity. What is the underlying rationale for combining Prototype, GVA, and IDAC? Beyond their individual performance improvements, is there any evidence that these modules provide complementary benefits?

  3. IDAC Constraints. The abstract mentions identity-preservation and normal/abnormal separation constraints, but Section 3.4 states that their loss weights are set to zero. Could the authors clarify which constraints are actually used in the reported experiments?

Reviewer_Q5skOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces Prototype–GVA–IDAC, an extension of MultiADS. MultiADS uses a labeled source dataset (containing tuples of image, anomaly label, anomaly mask) and CLIP to compute dense/patch embeddings of images and text embeddings of anomaly labels, and trains adapter heads on top of the patch embedding to maximize relative similarity to the correct anomaly type. To compute the final anomaly score it aggregates this dense scoring to the standard CLIP CLS to text embedding score.

This paper's main contributions lie in enhancing MultiADS with:

  • Prototype: Further contextualizing/refining the patch embeddings
  • GVA: Refining the image-level anomaly score using Prototype's output (and the CLS token)
  • IDAC: Adding calibrating learned offsets to the text embeddings

Improvements are reported on 5 datasets (BTAD, MPDD, KSDD2, VisA, and WFDD).

Critical strengths

None.

Critical weaknesses

The main weakness of this paper is that, from start to finish, its writing is extremely convoluted and hard to follow.

Furthermore, the method's modules are also convoluted, ad-hoc and critically, lacking any targeted ablations to support their specific design choices. For Prototype, the paper itself acknowledges this lack of ablations on "internal mechanisms".

Questions for authors

None

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

Related papers

Discussion (0)

Sign in to comment.