acceptodds
Under review as a conference paper at ICLR 2027

TRIAD: Calibrating Contextual and Normal-Support Evidence for Unsupervised Multi-Class Anomaly Detection

Abstract

Unified multi-class anomaly detection asks a single model, trained only on normal images, to detect and localize defects across many product categories. Scoring pipelines that inspect the target they are scoring suffer from two structural failures. First, attention lets target content inform its own score, so a plausible defect can be explained as normal. Second, raw error maps lack a common probabilistic scale across positions, appearance modes, and categories. We present TRIAD (Target-blind energy, Reference witness, and Integrated Anomaly Detection), a framework built entirely on a frozen DINOv2 backbone that computes three calibrated evidences for every test image. A target-blind contextual energy fills in masked patches from context alone and grades the observation by its likelihood under a von Mises–Fisher mixture over a library of normal patch atoms. A witness branch measures the distance to a coverage-maximizing library of normal images, aggregates the distance with a CVaR tail readout, and converts the aggregate into route-conditional empirical -values with finite-sample false-alarm control. A calibrated fusion places both fields on one commensurable scale and combines them with equal weights, a statistic that is likelihood-ratio optimal for dense shifts under a Gaussian working model, while image-level decisions remain identical to the witness detector. On MVTec-AD and VisA under the unified protocol, TRIAD advances the state of the art on all seven reported metrics, reaching / pixel AUROC, / pixel AP, and / AUPRO.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multi-class unsupervised anomaly detection, conditional energy, normal support estimation

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_Bq7pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper presents TRIAD for unified multi-class anomaly detection with three calibrated evidences on a frozen DINOv2 backbone. A target-blind contextual energy grades the clean observation by its likelihood under a context-conditioned von Mises-Fisher mixture over normal patch atoms. A witness branch aggregates the distance to a coverage-maximizing library of normal images with a CVaR tail readout and converts the aggregate into route-conditional empirical p-values with finite-sample false-alarm control. A calibrated fusion combines the two spatial fields with equal weights while image-level decisions remain identical to the witness detector. On MVTec-AD and VisA the paper reports gains over the unified baselines on all seven metrics and runs preregistered single-factor ablations for each evidence.

Critical strengths

The calibrated fusion is a clear strength. The witness branch computes a distance field that the image score alone would discard. The fusion recovers this field as a second localization evidence. In Table 4 the joint field outperforms each field alone on all four pixel-level metrics on both datasets while image-level decisions remain identical to the witness detector.

Critical weaknesses

W1. Inference-side ablation of target blindness. Table 2 tests target blindness, the first contribution, only at inference: the leaky variant feeds clean memory, an unmasked target, and a query containing the target projection into a frozen decoder trained with masked context. The drop could reflect this mismatch between training and test inputs rather than the proposed mechanism in which leakage lets the query explain the anomaly as normal. The paper does not report how energies at anomalous and normal positions change under this intervention. The shuffled variant tests context correspondence and cannot separate the two explanations. A leaky-trained counterpart is left to future work. Without it, Table 2 does not establish that target leakage causes the drop.

W2. Detection power of route-conditional conformal calibration. The finite-sample false-alarm guarantee is part of the second contribution, but Section 4.3 evaluates the image decision through false-positive-rate deviations without reporting detection rates at α′=0.05\alpha'=0.05. By Eq. (12), the p-value is bounded below by 1/(Nc(x)+1)1/(N_{c(x)}+1), so routes with fewer than 19 calibration images cannot flag any image. Toothbrush has 60 normal training images [1]. A 10% per-category split would give six calibration images and a minimum p-value of 1/7≈0.1431/7\approx0.143, consistent with the p^\hat{p} of its highest-scoring anomalous image in Figure 7. Such a route flags no image, and the false-alarm guarantee holds trivially. Figure 5 shows only pooled curves. The actual per-route calibration counts and per-category detection rates are needed.

W3. Incorrect author information in the references. The DiAD reference incorrectly includes Zhisheng Chen at line 540 and Yong Liu at line 541. Neither is an author of the published AAAI paper.

References [1] Bergmann et al. The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection. IJCV, 2021. Table 1.

Questions for authors

Q1. Missing comparison with Dinomaly. Table 1 and Section 2 omit Dinomaly [2], a DINOv2-based reconstruction method under the multi-class setting. Its default model reports higher values than TRIAD on five of the seven metrics on each dataset, e.g., 98.7 vs. 95.06 image AUROC and 53.2 vs. 46.61 pixel AP on VisA. Could the authors add it to Table 1, clarify any protocol differences, and revise the claim of state-of-the-art results on all seven metrics? Dinomaly also uses the mean of the top 1% pixels as the image anomaly score, the same top-tail averaging operation used in Eq. (11), although applied to a different anomaly field. Its Noisy Bottleneck and Unfocused Linear Attention aim to mitigate identity mapping. What additional benefit does target-blind contextual energy provide over these designs?

Q2. Calibration validity on MVTec-AD. Appendix E states that normal-image p-values follow the uniform reference, yet in Figure 5 (MVTec-AD) their empirical CDF appears above it over much of the range, indicating empirical false-positive rates above the nominal levels at those thresholds. Section 4.3 also reports a mean per-category deviation of 0.070 from the nominal level. If this denotes the mean absolute deviation from 0.05, at least one observed category-level false-positive rate must exceed 0.10. Could the authors clarify this statistic and report per-category false-positive rates at α′=0.05\alpha'=0.05, together with normal test counts and uncertainty estimates? Please also discuss whether the observed deviations can be explained by finite calibration and test samples, and whether the calibration images are exchangeable with the normal test images.

References [2] Guo et al. Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection. CVPR, 2025.

Reviewer_6yxrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies unified multi-class industrial anomaly detection using only normal images for training. A shared model detects anomalous images and localizes defects across product categories. The authors argue that existing methods can use the target content to explain away anomalies, while raw scores are not directly comparable across positions, appearance modes, and categories. They propose TRIAD, a framework built on frozen DINOv2 features.

The method has two base branches. The contextual energy branch masks target patches and uses a shallow decoder to predict mixture weights over a normal atom bank from the surrounding context. It then scores the observed features using an energy derived from a von Mises–Fisher mixture. The witness branch performs nearest-neighbour matching against a normal reference library. A CVaR tail readout of the witness distance field produces an image score, which is calibrated using route-conditional empirical p-values. The two spatial evidence fields are standardized and fused with equal weights for localization. Image-level decisions remain determined by the witness branch alone.

The paper also relates contextual energy to conditional negative log-likelihood and shows that equal-weight fusion has the same ranking as the optimal likelihood-ratio test under a restricted Gaussian working model. Experiments and branch ablations on MVTec-AD and VisA support the complementarity of the two evidence fields and show improved average localization performance. The main contribution is the combination of target-blind contextual energy, normal-support witness evidence, and score calibration in a unified detection framework.

Critical strengths
  1. Contextual energy and normal-support witness evidence complement each other and improve localization The paper combines two distinct sources of anomaly evidence:
  • Contextual energy: The target patch is masked, and the surrounding context predicts a normal feature distribution. The observed feature is then scored against that distribution. This asks whether the appearance at a position is consistent with its surroundings.
  • Normal-support witness evidence: A test patch is matched to the normal reference library using nearest-neighbour search. This asks whether the appearance is supported by similar normal features. Table 4 shows that fusing the standardized evidence fields improves pixel AUROC, AP, and AUPRO over the witness field alone. Pixel AP increases by 3.35 and 3.62 percentage points on the two datasets, respectively. This directly supports the complementarity of the two sources of evidence. Image-level decisions continue to use only the witness branch and are unaffected by pixel-level fusion.
  1. Training, calibration, and inference are described in detail, with ablations supporting the main design choices The paper explains the separate roles of the normal training, validation, and calibration splits, as well as the construction of the normal atom bank, reference library, and decoder. The appendix includes hyperparameters, theoretical derivations, and complete algorithms. The experiments compare branches, tail readouts, reference library selection, and calibration and fusion choices, and report results over multiple seeds. These details help explain the components and provide a useful basis for reproduction.
Critical weaknesses

W1. Computational cost is high, and comparisons with the closest existing methods could be stronger Tables 2–4 provide substantial ablation evidence for the contextual input design, witness branch, and calibrated fusion. However, the external comparisons are mainly against reconstruction-based methods. Comparisons with PatchCore- or AnomalyDINO-style methods using the same DINOv2 features and comparable resolutions and reference budgets would better establish TRIAD's advantage over closely related approaches. This is a limitation of the external comparison, not a lack of evidence for the contributions of the internal components.

TRIAD has a high computational cost. Table 7 reports 8585.8 G FLOPs, approximately 223 times the 38.5 G FLOPs of MaskAD. This is a clear disadvantage. Hardware differences affect FPS and training time, but do not explain the large FLOPs gap. The small additional runtime of the witness branch does not offset the high computational cost of the full method.

W2. The scope of blindness needs clarification, fixed-threshold evaluation is incomplete, and the fusion claim is too strong

Appendix B.1 derives the conditional likelihood with a fixed atom bank, but the routing in Eq. (3) and the standardization parameters selected in Eq. (9) still depend on clean target content. There is a clear basis for the direct blindness of the contextual predictor. However, the claim that there is no target-information pathway in the entire scoring pipeline needs to be narrowed. This does not mean that using the observed feature to evaluate its likelihood is itself leakage.

The add-one rule has a theoretical basis when the scoring rule is frozen and route-conditional exchangeability holds. However, detection performance at the paper's p-value threshold of 0.05 remains insufficiently evaluated. The FPR deviation reported on page 8 is not explicitly defined, and calibration sample counts, false-positive rates, and detection rates are not fully reported for each route. The highest-scoring anomalous toothbrush image in Figure 7 still has p=0.143. This illustrates that high AUROC does not necessarily imply successful detection at the chosen threshold.

Table 4 supports branch complementarity and the use of route-scalar nulls, but does not directly compare fusion without standardization or sum versus max-combination. Appendix B.2 explicitly limits equal-weight optimality to a Gaussian working model, yet the main text states that “fusion can only improve localization.” In Tables 8 and 9, AP decreases for tile, zipper, and pcb3 relative to the contextual field alone. The statement therefore goes beyond the experimental evidence.

Questions for authors

Q1. How does TRIAD compare with matched nearest-neighbour baselines?

Please compare with a unified multi-class PatchCore- or AnomalyDINO-style baseline using the same DINOv2 features, resolution, normal data, and a comparable reference budget, without ground-truth test categories for routing. Report image AUROC, pixel AP/AUPRO, and FLOPs to establish the advantage over closely related methods missing from Table 1.

Q2. Does calibration sample size limit detection at the 0.05 p-value threshold?

C.3 and Algorithm 2 set α′=0.05\alpha'=0.05. Eq. (12) implies p^(x)≥1/(Nc(x)+1)\hat p(x)\geq1/(N_{c(x)}+1), where Nc(x)N_{c(x)} counts calibration images for the predicted route; thus Nc(x)<19N_{c(x)}<19 prevents detection at this threshold. Is this why Figure 7's highest-scoring anomalous toothbrush image has p=0.143, or is misrouting or score overlap responsible? Please report per-route sample counts, image-level FPR and detection rates with confidence intervals, and define the FPR deviation reported on page 8.

Q3. What is the scope of target blindness?

B.1 fixes the atom bank, but routing in Eq. (3) and mode standardization in Eq. (9) depend on clean target content. Does the blindness claim apply only to mixture-weight prediction given a fixed bank? Please examine whether routing or mode selection can lower anomaly scores as target content changes, or narrow the claim on page 2.

Q4. What directly supports standardization and equal-weight fusion?

Table 4 supports complementarity and compares route-scalar with per-position nulls. With branch outputs, alignment, and post-processing fixed, compare Eq. (15) with (i) directly adding raw contextual energy (Eq. (8)) and witness distance (Eq. (10)), and (ii) taking the maximum of standardized zVz_V and zGz_G. Report pixel AP/AUPRO to support the sum-versus-max claim on page 6. Given B.2's limitations and tile's AP decrease in Table 8, please restrict “fusion can only improve localization” to average performance.

Q5. Are the gains from multi-scale masking worth the cost?

H.4 reports nearly identical single-scale and multi-scale contextual pixel AUROC (97.41/97.39 versus 97.42), but not full-model results. Given the high cost in Table 7, please compare full single-scale and multi-scale TRIAD in pixel AP, AUPRO, and FLOPs, with other settings fixed, to show whether the gains justify the extra computation.

Reviewer_M81JOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TRIAD, a unified multi-class anomaly detection framework built on frozen DINOv2 features. It introduces target-blind contextual energy modeling to reduce reconstruction shortcuts, a normal-reference witness branch with conformal calibration, and a calibrated fusion strategy for anomaly localization. The method combines masked contextual prediction, nearest-neighbor reference matching, and statistical score calibration. Experiments on MVTec-AD and VisA report substantial improvements over existing unified anomaly detection methods.

Critical strengths
  1. Interesting motivation and problem formulation. The paper investigates two meaningful challenges in unified anomaly detection: reconstruction shortcuts and inconsistent anomaly-score distributions across categories. The proposed contextual and reference-based evidence combination is conceptually reasonable.

  2. Strong reported experimental performance. TRIAD achieves competitive results on MVTec-AD and VisA, particularly for anomaly localization. The paper also provides detailed ablation studies and per-category evaluations supporting the effectiveness of its components.

Critical weaknesses
  1. Excessive computational complexity and questionable practical efficiency. Appendix Table 7 shows that TRIAD requires 8585.8 GFLOPs and achieves only 4.5 FPS, compared with 38.5 GFLOPs for MaskAD. Such substantial computational overhead raises serious concerns about practical industrial deployment. The authors should better justify this cost and evaluate more efficient alternatives.

  2. Limited independent contribution of contextual energy modeling. Table 4 shows that the witness branch alone already achieves 98.25/98.81 Pixel AUROC on MVTec-AD/VisA, while the complete TRIAD improves these results to 98.49/98.97. Since the expensive contextual branch provides relatively modest additional AUROC gains, its necessity and cost-effectiveness require stronger justification.

  3. Insufficient evidence for the target-blindness mechanism. The target-blindness ablation compares a trained blind decoder with inference-time leaky variants, rather than independently trained alternatives. Such comparisons may introduce distribution mismatch. Moreover, category routing and mode assignment still use features extracted from the unmasked query. The authors should clarify the precise scope of the claimed target blindness and provide fairer controlled comparisons.

  4. Questions about the unified learning paradigm. Although TRIAD employs a shared frozen backbone and decoder, it maintains category-specific atom banks, reference libraries, mode anchors, and calibration statistics. This raises questions about scalability and whether the method fully achieves the intended unified detection setting. The authors should report the complete memory overhead and evaluate robustness to category-routing errors.

  5. Incomplete comparisons and limited generalization. The evaluation is restricted to MVTec-AD and VisA, and the main comparisons largely rely on previously reported results. Stronger comparisons with DINO-based nearest-neighbor methods under matched feature extraction and computational budgets would help isolate the actual benefits of the proposed framework.

  6. Missing relevant recent work. The related-work discussion overlooks UniMMAD: Unified Multi-Modal and Multi-Class Anomaly Detection via MoE-Driven Feature Decompression (CVPR 2026). Although UniMMAD addresses a broader multi-modal setting, it is relevant to unified anomaly detection, category-specific adaptation, and efficient deployment. The authors should better position TRIAD relative to recent unified detection frameworks.

Questions for authors

See the weaknesses part.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.