acceptodds
Withdrawn from ICLR 2027

ProtoDeCo: Decoupling Anomaly Ranking and Normality Measurement for Cross-Domain Zero-Shot Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-domain zero-shot anomaly detection must rank target images without target-domain anomaly annotations. However, the representation that supports anomaly ranking need not provide a suitable metric for measuring target normality. We present ProtoDeCo, a post-hoc transductive calibration method that separates these roles: a fixed host detector selects low-score target images as pseudo-normal candidates, while frozen Contrastive Language-Image Pre-training (CLIP) patch features construct category-specific prototypes and measure deviations from them. The deviation score is fused with the host image score; the host pixel map remains unchanged. The method uses no target labels or masks and updates no model parameters. In the matched comparison, Detector Proto and Frozen Proto use the same host ranking, pseudo-normal candidates, memory construction, query reduction, and score fusion; they differ only in the representation used for metric measurement. Across 12 industrial host–target pairs, Frozen Proto improves the uncalibrated host by an average of 0.88 area under the receiver operating characteristic curve (AUROC) points, with positive gains on all pairs. On eight distinct-stream pairs, it improves over Detector Proto by 0.23 points; the 95% confidence interval is 0.14 to 0.36 points. On four medical image-level targets, applying ProtoDeCo to the uncalibrated BD-Host increases mean AUROC and average precision (AP) from 80.53 and 85.57 to 81.17 and 85.85, with gains on three targets. These results support separating anomaly ranking from target-side metric measurement under the evaluated transductive protocol.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical Anomaly Detection, Prototype-Based Measurement, Cross-Domain Generalization, Decoupled Representation Learning

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_8K19OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

ProtoDeCo separates anomaly ranking from target-side normality measurement. A fixed detector selects low-scoring target images, while frozen CLIP features form category-specific normal prototypes. Prototype distances calibrate image scores without changing pixel predictions. Matched feature-space comparisons and transductive controls evaluate the approach across industrial and medical datasets.

Critical strengths
  1. The matched Detector Proto/Frozen Proto comparison isolates the representation choice while holding candidate selection, memory construction, scoring, and fusion fixed. This provides useful evidence beyond comparisons between complete systems.

  2. The self-inclusion, cross-fitting, random-candidate, and target-batch controls address relevant transductive confounds. The paper also reports negative cases and clearly limits its contribution to image-level scoring.

Critical weaknesses
  1. The deployment setting limits practical applicability. ProtoDeCo requires a collection of unlabeled target images grouped by category to construct normal prototypes and normalize scores. Its predictions therefore depend on the available target batch, making applicability to online or incremental inspection unclear. It also requires post-hoc re-ranking of an existing detector’s outputs, adding a batch-dependent calibration stage to the deployment pipeline.

  2. The methodological novelty is limited. Pseudo-normal selection, frozen feature extraction, prototype clustering, nearest-neighbor scoring, and score fusion are common components in anomaly detection. The main contribution is to separate the representation used for anomaly ranking from that used for normality measurement. This is a reasonable design choice, but the method introduces no new distance-learning mechanism, prototype optimization algorithm, or theoretical account of target-domain adaptation. The contribution therefore appears primarily to be a recombination of established techniques.

  3. The practical gains remain modest relative to the added cost. Frozen Proto improves over Detector Proto by 0.23 AUROC points across eight distinct-stream pairs, and calibration adds 0.20 points to the main industrial host, while score-generation time increases by approximately 52%. Table 23 also reports stronger MuSc performance on MVTec/VisA and essentially no benefit from attaching the calibrator to MuSc. The advantage over existing transductive approaches therefore remains unclear.

Questions for authors

See Critical Weaknesses.

Reviewer_cx4cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ProtoDeCo, a post-hoc transductive calibration framework for cross-domain zero-shot anomaly detection. The method decouples anomaly ranking from normality measurement by using a fixed host detector to select pseudo-normal target images and frozen CLIP features to construct category-specific prototypes. The prototype-based deviation scores are then fused with the original detector scores without updating model parameters. Experiments on industrial and medical datasets demonstrate improvements in image-level anomaly detection performance.

Critical strengths

S1. The proposed decoupling of anomaly ranking and normality measurement provides an interesting perspective on cross-domain zero-shot anomaly detection.

S2. The method can be applied to pretrained anomaly detectors without additional training or parameter updates, making it practically appealing as a post-hoc calibration approach.

S3. The paper includes extensive experiments across industrial and medical datasets. The matched comparisons between Detector Proto and Frozen Proto, along with the ablation studies, provide useful evidence supporting the proposed design.

Critical weaknesses

W1. The performance improvements introduced by ProtoDeCo are relatively modest compared with its additional computational overhead. According to Table 3, applying ProtoDeCo to BD-Host improves the average industrial AUROC from 94.72% to 94.92%, while the average medical AUROC increases from 80.53% to 81.17%. However, Section 4.5 reports that the additional frozen feature pathway increases inference time from 233.59 ms to 354.72 ms per image. Although the additional computation does not require parameter updates, this overhead may limit the practical efficiency of the proposed calibration method. A more detailed discussion of the trade-off between accuracy improvements and computational costs would strengthen the practical contribution.

W2. The proposed method relies on the availability of an unlabeled target collection and category identifiers for pseudo-normal selection, prototype construction, and score normalization. This transductive setting may limit its applicability to practical scenarios where target images arrive sequentially or only small batches are available. Although the paper includes target-batch stress tests, Table 6 shows that the average improvement becomes negative when the batch size is reduced to eight images, with a worst-case AUROC degradation of 6.25. These results raise concerns about the robustness of the calibration method under limited target data. Further evaluation under streaming or incremental deployment settings would help establish its practical applicability.

W3. Although the paper provides hyperparameter sensitivity analyses in Appendix Table 20, the results suggest that the preferred fusion weight may vary across target datasets. For example, increasing the fusion weight improves performance on MVTec-AD but degrades performance on BTAD. Since the proposed method adopts a fixed fusion weight across different target domains and host detectors, further evaluation of its robustness across a broader range of datasets and detectors would strengthen its practical applicability.

Questions for authors
  1. Given the relatively modest performance improvements and the additional inference overhead, how do the authors justify the practical efficiency of ProtoDeCo? Could the authors provide further analysis of the accuracy–efficiency trade-off?

  2. Can ProtoDeCo be extended to streaming or incremental anomaly detection scenarios where the complete unlabeled target dataset is unavailable? How would the proposed calibration mechanism operate with only a small number of target images?

  3. Appendix Table 20 shows different performance trends with respect to the fusion weight across MVTec-AD and BTAD. How robust is the fixed fusion weight across different target domains and host detectors? Could the authors provide a more comprehensive sensitivity analysis?

Reviewer_c7NnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses cross-domain zero-shot anomaly detection. It proposes ProtoDeCo: scoring the unlabeled target images and sorts them into pseudo-normal and abnormal candidates.

Critical strengths

Improvements in Sec. 4 demonstrate empirical success.

Critical weaknesses
  1. The method cannot be understood from the main text because the key notation is undefined.
  • Several symbols that the method depends on are never defined: fi,uf_{i,u} and hi,uh_{i,u} (Eq. 4), UiU_i (Eq. 5), and "memory" (L191).
  • The notation πc(j)\pi_c(j) in Eq. 2 is only defined in the next equation, in a separate sentence. DjD_j, zijz_{ij}, and PcP_c are used in Fig. 1 before they are introduced.
  • "MiniBatchKMeans" (L191) also appears without explanation.
  • I had to guess the meaning of several of these from context, and I could not guess all.
  • I cannot verify the technical correctness of the approach.
  1. The contribution is not articulated, and the two main figures do not help.
  • L85 only says that ProtoDeCo uses several established components, each with a different role. This is not a contribution on its own, and the paper does not explain the purpose of each component or how the components differ from each other. I therefore cannot tell what is new.
  • Fig. 1 should convey the high-level idea, but the caption never explains subfigures (a) and (b), and the sentence after the L53 reference ("For each target category, the fixed host ranks ...") does not correspond to anything in the figure. I cannot tell whether (a) is a baseline and (b) the proposed method, whether they are two pipeline stages, or the same method in two scenarios.
  • Fig. 2 is referenced only once, in a single sentence at the end of Sec. 3, and has the same problem.
Questions for authors
  1. What is the role of MiniBatchKMeans at L191?
  2. In Fig. 1, how should (a) and (b) be read?
  3. What is the high-level role of each component in ProtoDeCo, and how do the components differ from each other?

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

Related papers

Discussion (0)

Sign in to comment.