acceptodds
Withdrawn from ICLR 2027

Supervised Anomaly Detection via Evidence-aware Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reconstruction-based anomaly detection methods have achieved strong performance by exploiting the discrepancy between reconstructed and observed patterns. However, reconstruction discrepancies are inherently ambiguous: large residuals may arise from difficult normal content, while anomalous regions can remain undetected when they are well reconstructed. This ambiguity leads to overlapping score distributions and limits reliable anomaly decisions under practical operating conditions. In this work, we propose Evidence-aware Reconstruction for Anomaly Detection (EviAD), a framework that uses a small set of annotated anomalies to augment reconstruction-based detectors with uncertainty-aware anomaly scoring. EviAD models two complementary forms of uncertainty: reconstruction uncertainty characterizes how difficult normal content is to reconstruct, helping reduce false alarms caused by large normal residuals; evidential uncertainty characterizes how strongly the learned evidence supports an anomaly prediction, helping identify anomalies that reconstruction errors alone may miss. Together, they guide anomaly scoring to better distinguish difficult normal samples from well-reconstructed anomalies. Extensive experiments demonstrate that EviAD consistently improves anomaly detection and localization performance, achieving higher recall under strict false-positive constraints and supporting more reliable industrial anomaly inspection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Anomaly Detection, Evidential Deep Learning, Uncertainty

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_6nzXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Evidence-aware Reconstruction for Anomaly Detection (EviAD), a supervised anomaly detection framework that augments reconstruction-based detectors with uncertainty-aware anomaly scoring. The method addresses the ambiguity of reconstruction errors by modeling two complementary uncertainties: reconstruction uncertainty for estimating normal reconstruction difficulty and evidential uncertainty for measuring the reliability of anomaly evidence. The framework consists of a reconstruction uncertainty module that predicts local reconstruction-error scales and refines reconstruction using difficulty-aware weighting, and an evidential uncertainty module that learns normal/anomaly reference subspaces and derives Dirichlet-based anomaly evidence. The final anomaly score combines calibrated anomaly belief with uncertainty-adjusted reconstruction responses. The authors evaluate EviAD on MVTec AD, Real-IAD, and MANTA under a K-anomaly support setting, reporting improvements over supervised and unsupervised anomaly detection baselines in image-level detection and pixel-level localization metrics.

Critical strengths
  1. The paper identifies an important limitation of reconstruction-based anomaly detection: reconstruction scores alone cannot distinguish difficult normal samples from true anomalies. This is a meaningful practical issue, especially for industrial inspection where false positives are costly.
  2. The experiments cover multiple industrial anomaly benchmarks (MVTec AD, Real-IAD, MANTA), include ablations, and evaluate operational metrics such as FPR@95 and Recall@1 rather than relying only on AUROC.
Critical weaknesses
  1. The central weakness is that EviAD appears to combine existing uncertainty estimation and prototype/evidence modeling techniques rather than introducing a new learning framework. The paper claims "Evidence-aware Reconstruction", but the novelty mainly lies in how existing components are assembled. For ICLR, the contribution may not reach the expected level of conceptual innovation.
  2. The experimental setup provides K annotated anomalies with pixel-level masks and synthetic augmentation. This makes the problem significantly easier than unsupervised anomaly detection. However, the paper compares against many unsupervised methods and highlights superiority.
Questions for authors

The proposed framework uses anomaly samples selected from the original test split as training supports. Could the authors clarify whether this follows previous supervised AD protocols and discuss the impact on practical deployment scenarios?

Since EUM relies on annotated anomaly masks, how does performance change when only image-level anomaly labels are available?

The reconstruction uncertainty module predicts reconstruction-error scales. How does this differ fundamentally from existing uncertainty-aware reconstruction methods or probabilistic reconstruction approaches?

Could the authors provide comparisons with more recent supervised anomaly detection methods beyond ArcAD, DRA, and BGAD?

How sensitive is EviAD to the number and diversity of anomaly supports? In particular, does the method still work when anomalies are highly heterogeneous?

Reviewer_9Y4dOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes EviAD, a supervised anomaly detection framework that leverages a limited number of labeled anomaly samples to improve reconstruction-based detection. It introduces reconstruction uncertainty modeling to reduce false positives in hard-normal regions and evidential uncertainty modeling to identify anomalies with low reconstruction errors. The latter learns easy-normal, hard-normal, and anomaly reference subspaces and employs Dirichlet evidence estimation for uncertainty-aware anomaly scoring.

Critical strengths
  • The paper pointed out a imporatant problem in practical industrial anomaly detection: the score overlap between normal and anomaly is large.
Critical weaknesses
  1. Limited technical novelty and insufficient distinction from existing works. Supervised anomaly detection with limited labeled anomalies has been extensively studied, and its core learning objective is closely related to binary classification or segmentation. So, the author should compare the performance with those methods. The proposed framework shares substantial similarities with existing methods, such as Catching Both Gray and Black Swans (CVPR 2022) and ArcAD (ECCV 2026).

  2. Limited Core Improvement The main technical contributions appear to be the EUM and RUM, including Dirichlet evidence estimation, and hard-normal refinement, which largely build upon established techniques. In the ablation study, the improvement of EUM and RUM is very samll.

  3. Insufficient validation of the claimed reliability and generalization. The paper emphasizes reliable anomaly detection under practical operating thresholds. However, it does not provide an explicit strategy for determining the final detection threshold. The reported FPR@95 and Recall@1 metrics demonstrate score discrimination at particular ROC operating points, but do not establish whether comparable performance can be achieved using thresholds determined without test labels. Furthermore, the paper lacks explicit evaluation on unseen anomaly types, which is particularly important when learning from a limited number of labeled anomalies. These limitations weaken the claimed practical significance and generalization capability of the proposed framework.

Questions for authors
  1. Necessity of the three-subspace design. The authors divide normal features into easy-normal and hard-normal groups, together with an anomaly subspace. However, it is unclear why three separate subspaces are necessary. Could the authors compare this design with a simpler two-subspace formulation (normal vs. anomaly) and explain the advantages of explicitly separating easy-normal and hard-normal features?

  2. Effectiveness of Dirichlet evidence modeling. The improvements of EUM may come from supervised feature discrimination rather than uncertainty estimation itself. Could the authors provide more targeted ablations comparing Dirichlet evidence with direct affinity scores, softmax probabilities, or simpler confidence-weighting strategies? Additionally, can the authors evaluate the quality of the estimated uncertainty using calibration metrics?

  3. Consistency between training and inference. In Section 3.3, the supervised loss is applied to the expected anomaly probability y^A=αA/Sα\hat{y}_ A = \alpha_A / S_\alpha, whereas inference uses the anomaly belief bA=eA/(S+2)b_A = e_A / (S+2). Could the authors explain the motivation for using different quantities during training and inference, and whether this discrepancy affects prediction calibration?

  4. Computational complexity and efficiency. The framework introduces additional uncertainty estimation, reference-subspace learning, and multiple training stages on top of a reconstruction backbone. Could the authors report the additional training and inference costs and discuss whether the performance improvements justify the increased complexity?

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

This paper presents EviAD, which improves reconstruction-based anomaly detection using a small number of anomaly images with pixel-level annotations. The method has two branches. One predicts the reconstruction error scale of normal regions and uses it to refine reconstruction and adjust anomaly scores. The other matches local features against normal and anomalous reference directions to provide additional anomaly evidence. Experiments on three industrial datasets show improvements in detection, localization, and recall at low false-positive rates.

Critical strengths

The paper evaluates anomaly detection under low false-positive constraints, rather than relying only on AUROC. On Real-IAD, recall at 1% FPR increases from 59.6% to 64.4% compared with the strongest supervised baseline, while FPR at 95% recall decreases from 28.1% to 24.8%. These results provide concrete evidence of practical value. The method improves performance on three industrial datasets and reports results under different selections of anomaly supports. The component experiments also show that combining reconstruction with supervised anomaly evidence works better than using the evidence branch alone. This supports the overall decision to combine these two sources of information.

Critical weaknesses

The results show that the full system works, but they do not directly establish the quality of its uncertainty estimates. Removing the entire evidence loss also removes the supervised segmentation terms. Replacing the evidence branch with a binary head changes the reference structure and training setup as well. These comparisons therefore do not isolate the contribution of uncertainty modeling. The paper should also explain more clearly why the matching-based evidence is suitable for the chosen Dirichlet formulation. The method uses predicted scales to guide an additional reconstruction refinement stage. However, it does not compare this with ordinary fine-tuning, residual-based hard mining, or focal weighting under the same training budget. Such a comparison would be particularly useful because the extra gain over reconstruction plus evidence is modest.

Questions for authors

Under the same training budget, does weighting by predicted scales outperform weighting by actual residuals or ordinary fine-tuning? Could you report inference correction and reconstruction refinement separately?

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

Related papers

Discussion (0)

Sign in to comment.