acceptodds
Under review as a conference paper at ICLR 2027

HIVE-AD: Adaptive Multilevel Feature Fusion for Training-Free Few-Shot Industrial Anomaly Detection

Abstract

Few-shot industrial anomaly detection requires reliable characterization of normal patterns from only a handful of defect-free reference images, making repeated optimization for new products impractical for rapid deployment. Existing training-free methods often combine features from different layers using fixed rules and rely on a shared anomaly response for image-level detection and pixel-level localization. These strategies can miss complementary texture and structural cues, magnify harmless variations in normal samples, and produce fragmented anomaly maps. To overcome these limitations, we introduce HIVE-AD, a training-free framework that adaptively fuses multilevel features from a frozen DINOv2 backbone. HIVE-AD constructs an independent normal memory for each selected layer, determines each layer's contribution from spatial variation in normal support images, and adopts separate scoring procedures for image-level ranking and pixel-level localization. The visual pipeline requires no parameter updates, prompt tuning, or additional training for individual product categories. An optional frozen MLLM further translates localized anomaly regions into structured defect reports. In the one-shot setting, HIVE-AD achieves image-level and pixel-level AUROC scores of 97.8% and 97.6% on MVTec-AD, and 94.5% and 98.2% on VisA, respectively. Compared with a frozen DINOv2-G baseline, HIVE-AD improves image-level AUROC by 2.3 and 2.8 percentage points on MVTec-AD and VisA, demonstrating that adaptive multilevel feature fusion and task-specific scoring improve few-shot anomaly detection without additional training.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Few-Shot Industrial Anomaly Detection, Training-Free Learning, Adaptive Multilevel Feature Fusion, DINOv2.

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_8jHoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes HIVE-AD, a training-free, few-shot industrial anomaly detection (IAD) framework built on a frozen DINOv2 backbone. The method addresses three design axes that the authors argue are underserved by prior training-free pipelines: (i) representation depth — Variance-Guided Layer Weighting (VGLW) computes a spatial-variance statistic over the normal support features of the deepest selected layer and deterministically maps it to one of three fixed layer-weight templates (early/balanced/late emphasis) over five selected DINOv2 blocks {11, 17, 23, 29, 39}; (ii) output granularity — Decoupled Pixel-Ranking Scoring (DPRS) separates image-level ranking (Gaussian smoothing + top-ratio pooling) from pixel-level scoring; (iii) spatial extent — Pixel-Level Soft Spatial Aggregation (PSSA) blends a smoothed anomaly map with a local max-filtered response to consolidate fragmented responses while remaining score-continuous. Patch anomaly evidence comes from layer-wise normal memory banks with exact cosine 1-NN matching. An optional frozen MLLM (InternVL2-8B) converts localized regions into structured defect reports, but is explicitly decoupled from all quantitative predictions. Validation covers MVTec-AD and VisA under 1/2/4-shot settings. In the 1-shot setting, HIVE-AD reports 97.8/99.0 I-AUROC/I-AUPR, 97.6 P-AUROC, 94.6 PRO on MVTec-AD, and 94.5/95.3, 98.2, 93.4 on VisA. It improves image-level AUROC over the bare DINOv2-G baseline by 2.3 and 2.8 points, ablations show each component contributes, and a sensitivity analysis (Table 3) shows VGLW is stable across threshold choices and a continuous interpolation. The supplement provides a single global configuration (Table A2), runtime/memory measurements, category-level results, a decision-threshold diagnostic (ZS vs. oracle accuracy), an automatic audit of the report branch (location/text-align/type accuracy), and a limitations discussion.

Critical strengths
  1. Rigorous and transparent evaluation protocol with a genuinely global configuration. All results (main paper + supplement) use one fixed hyperparameter set (Table A2) shared across datasets, categories, shot counts, and backbone scales, with no category-conditioned overrides. The VGLW sensitivity study (Table 3) directly probes the most heuristic component and shows I-AUROC varies by at most 0.45/0.54 points across six settings. The authors explicitly mark VisA PRO as a loss to SubspaceAD and explicitly state (B.7) that the comparison table is not a compute-matched backbone replacement study. This level of calibration transparency is above average for this subfield and materially strengthens the credibility of the headline numbers.
  2. Clean architectural separation between quantitative detection and language reporting. The MLLM branch is invoked only after scores, decisions, and maps are finalized, its output is neither fused into nor allowed to veto the visual result (B.5), and enabling/disabling it leaves all metrics unchanged. The automatic audit (Table A8) reports modest numbers (e.g., location 62.9, type 65.3 on MVTec-AD 1-shot) and the authors themselves caution that these are "diagnostic rather than evidence of fully faithful explanations." This honest treatment avoids the common failure of overclaiming MLLM-based anomaly explanation.
Critical weaknesses
  1. Limited conceptual novelty; each component is a straightforward heuristic, and the gains over the closest training-free baselines are modest. VGLW is a 3-way hard thresholding of hand-fixed weight templates; DPRS is Gaussian smoothing plus top-ratio pooling; PSSA is a max-filter blend. The 1-shot improvement over SubspaceAD is +0.7 I-AUROC on MVTec-AD and +1.3 on VisA (Sec. 4.2), while on VisA PRO HIVE-AD actually loses to SubspaceAD (93.4 vs. 95.5 in the 1-shot row of Table 1), and at 4-shot on VisA it also trails in I-AUROC (95.3 vs. 94.7 is a gain, but P-AUROC 98.6 vs. 98.4 and PRO 94.7 vs. 96.0 again show mixed results). The ablation (Table 2) shows the components are complementary, but the underlying mechanism of why the spatial-variance statistic should map monotonically to depth emphasis is only asserted (Sec. 3.2), not analyzed. For a paper whose core claim is a new way of organizing evidence, the conceptual depth behind the mapping is thin.
  2. Provenance of the global hyperparameters is unclear, which undermines the "no tuning" claim. The paper emphasizes that (τ_1,τ_2)=(0.8,1.5), (σ_r,r)=(4.0,0.002), (σ_p,h,α)=(8.0,31,0.6) form "one global configuration shared by both datasets" (Sec. 4.1). But it does not state how these values were chosen — in particular whether they were selected using the same test sets on which they are reported. Table 3 shows the default (0.8, 1.5) achieves "best or tied-best on four of six dataset-metric combinations," which reads as if the default was verified (or selected) against test performance. The same concern applies to the support-set choice for K-shot evaluation (no variance over random support draws is reported anywhere, despite single-support results like macaroni2 1-shot I-AUROC of 84.8 vs. 81.6 at 4-shot in Table A6 suggesting high support sensitivity). Additionally, the decision-threshold diagnostic (Table A3) reveals a large gap between zero-shot accuracy (82.9 / 73.2) and oracle accuracy (95.7 / 88.7) at 1-shot, indicating the CutPaste-style threshold calibration is far from effective — the binary-decision pathway is weak even though the paper correctly excludes it from the main claims.
Questions for authors
  1. Tables 1/A4/A5/A6 report point estimates. Could you report mean ± std over multiple random draws of the K-shot support set (or at least the minimum/maximum over draws) for the headline numbers? The large category-level swings (e.g., macaroni2, fryum) suggest support-set sensitivity may be significant at K = 1.
  2. In Eq. (3), v_s is computed without ℓ2 normalization from raw frozen features, yet the matching in Eq. (2) uses ℓ2-normalized features. What is the empirical scale/distribution of v_s across the 27 categories of MVTec-AD + VisA? Does the three-way template assignment actually vary across categories, or does nearly every category land in one bucket (which would reduce VGLW to a fixed weighting)?
  3. Sec. H attributes the VisA PRO gap to "under-coverage of large or fragmented anomalous regions." A simple diagnostic (e.g., PRO restricted to images with large GT-connected components, or coverage vs. PRO curves for varying α) would substantiate this claim. Could you provide it?
  4. Since DPRS already decouples ranking from localization, extending the localization branch with a lightweight multi-scale PSSA (as you suggest in Sec. H) seems the most direct route to closing the PRO gap; I would encourage this for a future version rather than requiring it now.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 8jHo (Q1) (Part I)

We thank the reviewer for the encouraging assessment, for recognizing our work on a unified evaluation configuration, experimental transparency, and the decoupling of visual detection from language reporting, and for raising questions about variation across support-set draws, spatial variance and template assignment, and anomaly region coverage on VisA.

Q1.

· Mean and standard deviation of overall metrics across random support sets

We thank the reviewer for suggesting that we report results across multiple random support sets. For each shot setting on both datasets, we use three random seeds to draw the required number of support images from the normal training images in each category. For each evaluation, we first compute the category macro-average and then report the mean ± standard deviation across the three runs, as shown below.

MVTec-AD

SettingI-AUROCI-AUPRP-AUROCPRO
1-shot97.69 ± 0.0798.98 ± 0.0997.61 ± 0.0694.77 ± 0.30
2-shot97.40 ± 0.1898.89 ± 0.1697.88 ± 0.0395.23 ± 0.72
4-shot98.31 ± 0.4399.44 ± 0.3198.10 ± 0.1795.74 ± 0.16

VisA

SettingI-AUROCI-AUPRP-AUROCPRO
1-shot94.63 ± 0.3095.14 ± 0.1698.18 ± 0.1393.59 ± 0.43
2-shot95.13 ± 0.5595.44 ± 0.1798.48 ± 0.0794.23 ± 0.27
4-shot95.68 ± 0.4695.32 ± 0.4398.66 ± 0.0394.65 ± 0.23

In the 1-shot setting, I-AUROC is 97.69 ± 0.07 on MVTec-AD and 94.63 ± 0.30 on VisA. Across the two datasets and three shot settings, the standard deviation ranges from 0.07 to 0.55 percentage points for I-AUROC and from 0.03 to 0.17 percentage points for P-AUROC. The largest standard deviation is 0.72 percentage points for 2-shot PRO on MVTec-AD. On both datasets, mean P-AUROC and PRO increase with the number of support images per category, suggesting that more normal reference images help the visual branch achieve better localization in repeated evaluations.

AuthorsOpenReview ↗

Response to Reviewer 8jHo (Q1) (Part II)

· Category-level results on MVTec-AD and VisA in the 1-shot, 2-shot, and 4-shot settings

MVTec-AD

Category1-shot I-AUROC1-shot P-AUROC2-shot I-AUROC2-shot P-AUROC4-shot I-AUROC4-shot P-AUROC
bottle99.9498.9799.9499.00100.0099.07
cable94.6596.5995.4797.0698.1997.56
capsule97.4898.1595.5898.2996.8598.52
carpet99.9499.3099.9899.31100.0099.34
grid99.9999.4099.9999.42100.0099.47
hazelnut99.0099.4099.5999.46100.0099.52
leather99.9999.2599.9999.25100.0099.22
metal_nut99.9296.2999.9997.10100.0097.55
pill96.7695.0296.3695.4497.2595.87
screw81.4397.9077.2897.6384.0998.29
tile100.0097.5699.9997.77100.0097.74
toothbrush99.9198.6299.8098.86100.0098.98
transistor96.8993.1997.7494.9198.3995.65
wood99.5696.7199.4296.9999.8896.86
zipper99.8997.8099.8897.71100.0097.86
Mean97.69 ± 0.0797.61 ± 0.0697.40 ± 0.1897.88 ± 0.0398.31 ± 0.4398.10 ± 0.17

VisA

Category1-shot I-AUROC1-shot P-AUROC2-shot I-AUROC2-shot P-AUROC4-shot I-AUROC4-shot P-AUROC
candle93.5599.3994.3299.4694.3299.49
capsules96.1498.2496.7698.8097.7599.04
cashew95.6897.4496.2997.9596.9398.15
chewinggum99.2599.6599.5199.6399.8599.60
fryum96.2093.8695.8194.4196.3594.76
macaroni190.1698.8092.8699.0392.4599.13
macaroni286.2998.3584.9798.7086.3699.31
pcb194.8899.1693.3999.2393.6099.28
pcb291.5998.3292.2898.8093.5998.97
pcb394.3798.4496.8398.8297.9299.01
pcb498.7097.6898.9997.9699.6898.16
pipe_fryum98.7598.8399.5598.9799.3699.02
Mean94.63 ± 0.3098.18 ± 0.1395.13 ± 0.5598.48 ± 0.0795.68 ± 0.4698.66 ± 0.03
AuthorsOpenReview ↗

Response to Reviewer 8jHo (Q2)

Q2.

· Different normalization procedures in Eqs. (3) and (2)

We thank the reviewer for drawing attention to the different normalization procedures used for spatial variance and cosine matching. We compute vsv_s in Eq. (3) before applying the additional patch-wise ℓ2\ell_2 normalization. This statistic measures the variance of the normal support features from the deepest selected layer across spatial positions and then averages it over support images and channels. It describes spatial variation within the normal reference images. The ℓ2\ell_2 normalization in Eq. (2) serves cosine-distance matching, so the anomaly score depends on differences in feature direction. The two computations provide the signal for layer-weight selection and the normalized features for matching, respectively.

We further compare the category rankings before and after normalization across three support-set draws. The rank correlation between the category means of the original vsv_s and those of the variance after patch-wise ℓ2\ell_2 normalization is ρ=0.950\rho=0.950 on MVTec-AD and ρ=0.965\rho=0.965 on VisA. This additional diagnostic shows that the relative category rankings are highly consistent under the two statistics, providing further evidence for selecting layer weights based on spatial variation in normal support features.

· Empirical scale and threshold intervals of vsv_s

We thank the reviewer for asking about the numerical scale of vsv_s and its relationship to the templates. We use the same pair of thresholds, 0.8 and 1.5, throughout. The early template corresponds to vs<0.8v_s<0.8, the uniform template to 0.8≤vs≤1.50.8\leq v_s\leq1.5, and the late template to vs>1.5v_s>1.5. Among the categories examined across three draws, cashew, carpet, grid, leather, tile, and wood consistently fall in the middle interval, and chewinggum and macaroni2 consistently fall in the upper interval. For screw, vsv_s falls in the upper interval in all three draws, and the selected late template achieves the highest I-AUROC among the candidates in each comparison. These empirical intervals directly link the observed template assignments to the scale of the statistic. The revised manuscript will report vsv_s and the template assignment for each category using the same procedure, so readers can see how categories from both datasets are distributed across the threshold intervals.

· Variation in template assignment across categories and support sets

We thank the reviewer for suggesting that we check whether the three templates actually vary across categories. In the three draws used for this assignment diagnostic, several categories consistently select uniform, while chewinggum and macaroni2 consistently select late. These assignments illustrate differences in weighting preferences across categories. Once the support features determine the template, its weights remain fixed for all query images associated with that support set. This preserves adaptability to the support set in few-shot settings and ensures consistent evaluation conditions within each support set.

AuthorsOpenReview ↗

Response to Reviewer 8jHo (Q3)

Q3.

· Diagnostic setup for examining insufficient region coverage on VisA

We thank the reviewer for suggesting a coverage diagnostic based on anomaly region shape. On VisA, we use three random seeds to draw support sets. For each seed, we keep the support set, query set, and other parameters fixed and vary the PSSA mixing weight α over 0, 0.2, 0.4, 0.6, and 0.8. We report coverage for large regions, regions with multiple connected components, and regions with a single connected component based on the size and connectivity of the ground-truth regions. The multi-component group is used to examine fragmented anomalies. We compute all coverage values at 5% FPR and report the mean across the three random seeds. A fixed false-positive operating point allows direct comparison of coverage at different values of α and shows which types of anomaly regions benefit.

αLarge-Region Coverage (%)Multi-Component Region Coverage (%)Single-Component Region Coverage (%)
088.2685.1594.65
0.289.5286.4794.90
0.490.2287.1094.86
0.690.5487.2394.54
0.890.5586.9694.06

At α = 0, coverage for large and multi-component regions is 88.26% and 85.15%, respectively, which is 6.39 and 9.50 percentage points lower than coverage for single-component regions. As α increases from 0 to 0.6, coverage improves steadily for both groups. At the default setting of α = 0.6, coverage increases by 2.28 percentage points for large regions and by 2.08 percentage points for multi-component regions, while single-component coverage decreases by only 0.11 percentage points. These results show that the main benefit of PSSA is to reinforce the responses in large and fragmented anomaly regions, improving coverage for these groups at the same false positive rate. The results for the individual groups support the explanation in Section H that region coverage may affect PRO on VisA.

Reviewer_Q6ZtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes HIVE-AD, a training-free framework for few-shot industrial anomaly detection using frozen DINOv2 features. It constructs separate normal patch memories at five transformer layers and computes cosine nearest-neighbor distances. Variance-Guided Layer Weighting (VGLW) selects one of three predefined weight templates based on spatial variation in the deepest support features. Decoupled Pixel-Ranking Scoring (DPRS) separates image-level ranking from localization, while Pixel-Level Soft Spatial Aggregation (PSSA) refines the localization map through Gaussian smoothing and local maximum filtering. An optional frozen InternVL2-8B generates structured defect reports without changing the visual predictions.

Experiments cover MVTec-AD and VisA under one-, two-, and four-shot settings. In the one-shot setting, the reported image/pixel AUROCs are 97.8/97.6 on MVTec-AD and 94.5/98.2 on VisA. The paper also provides component ablations, backbone scaling, category-level results, threshold diagnostics, and reporting analyses.

Critical strengths
  1. A well-motivated design for few-shot adaptation. Avoiding category-specific optimization is useful when new products have only a few normal reference images. HIVE-AD combines straightforward operations into a complete detection and localization pipeline, with a single configuration shared across both benchmarks, categories, and shot settings. This reduces the need for product-specific intervention.

  2. A promising support-conditioned fusion idea. VGLW uses statistics available from the normal support images to adjust the emphasis across feature depths. The rule is inexpensive to compute, requires no anomaly labels or parameter updates, and can operate with one support image. Adapting the use of frozen representations to the observed product appearance is a sensible direction, even though the proposed statistic still needs more direct validation.

  3. Competitive results with a useful matched baseline. The one-shot image AUROCs of 97.8 on MVTec-AD and 94.5 on VisA are competitive. Relative to the internal baseline, the reported gains are 2.3 and 2.8 percentage points. Section 4.3 matches the DINOv2-G backbone, input resolution, feature layers, and exact cosine 1-NN matching, which makes this comparison more informative than a backbone upgrade alone. The ablation inconsistency discussed below limits component-level attribution but does not, by itself, invalidate the full-system comparison.

  4. Substantial and relevant experimental coverage. The experiments cover two benchmarks, 27 categories, and three support sizes. Single-component and combination ablations examine the pipeline design; threshold and mapping sensitivity studies investigate the weighting heuristic; backbone scaling explores dependence on model size; and category-level results expose variation hidden by aggregate scores. These studies address several relevant questions and represent a meaningful strength of the submission, although they do not yet resolve every attribution or reliability issue.

  5. A complete system with clearly separated claims. Runtime and memory measurements, threshold diagnostics, and a preliminary report audit extend the evaluation beyond benchmark accuracy. The optional MLLM can be disabled without changing visual predictions. Oracle accuracy is explicitly identified as diagnostic, and the authors acknowledge weaker VisA PRO and limited report faithfulness. This transparency makes the scope of the results easier to assess.

Critical weaknesses

W1. Component attribution requires clarification of an ablation inconsistency

According to Eqs. (5)–(9), the image score depends on the weighted coarse anomaly map, ranking-branch smoothing, and top-ratio pooling. PSSA operates only on the localization branch. Therefore, with DPRS enabled and the support set, features, weights, and ranking parameters fixed, adding PSSA should leave every image score unchanged.

Table 2 nevertheless reports the following changes:

ComparisonDatasetImage AUROCImage AUPR
DPRS → DPRS + PSSAMVTec-AD96.2 → 97.597.7 → 98.5
DPRS → DPRS + PSSAVisA93.4 → 94.593.4 → 95.0
VGLW + DPRS → FullVisA93.8 → 94.594.7 → 95.3

These comparisons already enable DPRS, so a different definition of the coupled baseline does not explain the discrepancy. The authors should clarify whether there is an undocumented dependency, whether other experimental conditions change between rows, or whether the table contains reporting errors. If different support sets were used, the comparisons would not isolate PSSA's contribution. This is an important uncertainty in interpreting the component gains. It may be resolvable through precise configuration descriptions and matched evaluations, and it does not by itself establish that the complete detector's reported performance is invalid.

W2. The distinctive value of adaptive weighting needs a more direct comparison

HIVE-AD's contribution is best understood as a task-specific integration of support-conditioned fusion and carefully designed readouts. AnomalyDINO already computes image scores by averaging the highest patch distances while obtaining localization maps through a separate upsampling and Gaussian-smoothing path. HIVE-AD makes concrete changes to pooling resolution, smoothing order, and pooling ratio, but should position these changes more precisely relative to that existing separation. AnomalyDINO, arXiv:2405.14529v3, Section 3.1 and Appendix C.2

VGLW is the more distinctive design, and Table 3 usefully shows limited aggregate sensitivity to several thresholds and mappings. What remains unclear is why larger spatial variance in the deepest support features predicts a preference for deeper layers. Spatial variance can also reflect background composition, object occupancy, and feature magnitude. The existing sensitivity study and uniform-weight control do not establish whether support-dependent routing improves over using an early- or late-biased template globally.

The most informative comparison would hold the entire remaining pipeline and support sets fixed while evaluating fixed-early, uniform, fixed-late, and VGLW weighting. Category-level differences and routing frequencies would help explain when the proposed statistic is useful. The choice to compute variance from unnormalized features also warrants clarification, especially when the same thresholds are applied across backbone sizes.

W3. The statistical reliability and parameter-selection protocol are underspecified

The manuscript does not clearly specify support-image selection, independent repetition counts, random seeds, or whether the ablations share identical support sets. The main results and ablations lack uncertainty estimates. Determinism conditional on a support set does not remove variation caused by choosing that set.

This matters for the reported one-shot image-AUROC gains of 0.7 and 1.3 percentage points over SubspaceAD, and particularly for small pixel-level differences. SubspaceAD's revised Table 1 reports five-seed means and standard deviations, including 97.1 ± 0.9 and 93.2 ± 0.8 for one-shot image AUROC. HIVE-AD cites the corresponding means correctly, but its own sampling uncertainty is needed to assess the comparison. These baseline deviations do not prove that HIVE-AD's gains are insignificant. SubspaceAD, arXiv:2602.23013v3, Table 1 and Section 4.3; identified through Paper Lantern and checked against the primary source

Using one global configuration is a strength, but the paper should explain how the selected layers, weighting templates, thresholds, pooling ratio, and spatial-filter parameters were chosen, including which data informed those choices. The present description does not establish test-set tuning, but leaves the selection protocol unclear. The authors appropriately acknowledge that Table 1 compares complete published systems rather than fully matched compute configurations; consequently, that table alone cannot isolate the algorithmic contribution.

W4. Localization evaluation and numerical aggregation need clarification

Equation (9) resizes the final anomaly map to 224×224, despite extracting features from 672×672 inputs. The manuscript should specify the metric evaluation resolution, ground-truth mask interpolation, PRO integration range, and aggregation across images and categories. These choices are consequential for small defects and for a method whose main refinement changes spatial coverage.

There is also a directly checkable reporting discrepancy. The 15 displayed one-shot category P-AUROCs in Table A5 average to approximately 97.713, whereas its Mean row reports 97.6 and Table 3 reports 97.64 for the default setting. Ordinary rounding of the displayed category values does not reconcile these numbers under equal-weight averaging of the same runs. Please clarify the aggregation or correct the table.

Finally, the one-shot VisA PRO of 93.4 is below SubspaceAD's 95.5 despite equal reported pixel AUROC. The authors acknowledge this limitation, but the proposed explanation of under-coverage remains a hypothesis. A focused analysis of coverage errors would strengthen the interpretation of PSSA's effect.

Questions for authors
  1. Ablation consistency (W1): When PSSA is toggled with DPRS already enabled, are the per-image scores identical? Please provide the exact configurations and explain the changes in Table 2 using matched support sets and ranking parameters.

  2. Adaptive weighting (W2): Does VGLW outperform each of the three fixed templates under paired support-set evaluations? How do the observed support variances and template choices relate to category-level improvements?

  3. Statistical and selection protocol (W3): How many independent support sets were evaluated, how were they selected, and were they shared across ablations? Please report uncertainty and explain the data and procedure used to select the global hyperparameters.

  4. Localization evaluation (W4): At what resolution are pixel metrics computed, how are masks resized, and how is PRO integrated and aggregated? What explains the Table A5 mean discrepancy?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Q6Zt (Q1)

Q1.

· Exact configurations for matched ablations

We thank the reviewer for asking us to specify the paired ablation configurations. The comparisons use the same data splits and support/query samples, with a frozen DINOv2-G backbone, 672 × 672 inputs, and feature blocks [11, 17, 23, 29, 39]. In each pair, the DPRS image-ranking branch uses σᵣ = 4.0 and r = 0.002. We make separate paired comparisons for uniform and VGLW weights. With PSSA enabled, the localization branch uses σₚ = 8.0, h = 31, and α = 0.6. Both rows in each comparison share the same layer weights and ranking settings, so the per-image scores are compared under the same image-ranking procedure.

· Image-level metrics and localization gains in Table 2

We thank the reviewer for identifying the inconsistency in Table 2. The differences in the original table arose from an error in how we recorded the image-level metrics. To verify the reviewer’s observation, we kept DPRS enabled, fixed the support set, query set, layer weights, and DPRS ranking parameters, and toggled only PSSA. We compared the raw anomaly score for each query image with PSSA enabled and disabled. The maximum absolute difference across the paired scores was exactly zero ((\delta_{\max}=0)). This confirms the reviewer’s expectation that toggling PSSA leaves every image-level anomaly score unchanged under matched conditions.

On MVTec-AD, both “+ DPRS” and “DPRS + PSSA” achieve an I-AUROC of 96.2% and an I-AUPR of 97.7%. On VisA, both metrics are 93.4% for both configurations. With VGLW weights, the image-level metrics for “VGLW + DPRS” likewise match those of the full configuration. I-AUROC and I-AUPR are 97.8% and 99.0% on MVTec-AD and 94.5% and 95.3% on VisA, respectively.

With uniform weights, adding PSSA to the localization branch improves P-AUROC and PRO by 1.200 and 1.600 percentage points on MVTec-AD and by 0.200 and 0.800 percentage points on VisA, respectively. With VGLW weights, the corresponding increases are 0.300 and 0.700 percentage points on MVTec-AD and 0.400 and 1.000 percentage points on VisA. These paired results show that PSSA improves localization while preserving the per-image anomaly scores. We have corrected the affected entries in Table 2 and incorporated these results into the main text of the revised manuscript. MVTec-AD

VariantI-AUROCI-AUPRP-AUROCPRO
+ DPRS96.297.796.392.4
VGLW + DPRS97.899.097.393.9
Full CAFE97.899.097.694.6

VisA

VariantI-AUROCI-AUPRP-AUROCPRO
+ DPRS93.493.497.692.5
VGLW + DPRS94.595.397.892.4
Full CAFE94.595.398.293.4
AuthorsOpenReview ↗

Response to Reviewer Q6Zt (Q2)

Q2.

· Paired performance comparison between VGLW and three fixed templates

We thank the reviewer for asking for a separate comparison with each fixed template on the same support sets. We compare VGLW with early, uniform, and late using the same three 1-shot support-set draws and keeping all other settings fixed. Each entry reports the mean paired difference between VGLW and the corresponding fixed template in percentage points. All 12 differences across the four dataset and metric combinations are positive. Compared with the best fixed template for each combination, VGLW achieves mean gains of 0.02 to 0.22 percentage points. For pixel-level AUROC on MVTec-AD, VGLW improves on uniform, the best fixed template, by 0.03 percentage points. For image-level AUROC on VisA, VGLW improves on late, the best fixed template, by 0.22 percentage points.

Dataset / Metricvs. Earlyvs. Uniformvs. Late
MVTec-AD i-AUROC+3.23+0.68+0.02
MVTec-AD p-AUROC+0.41+0.03+0.09
VisA i-AUROC+3.20+0.25+0.22
VisA p-AUROC+0.89+0.10+0.02

· Relationship between support spatial variance, template selection, and category-level gains

We thank the reviewer for asking how support feature statistics relate to category-level gains. VGLW uses the spatial variance of normal support features to choose among three depth-weighting configurations. In the category-level case analysis, wood uses uniform and screw uses late, and each achieves the best result among the candidates in all three comparisons. For screw, late improves mean image-level AUROC by 9.339 percentage points over uniform. The two product categories achieve their respective best results with different weight configurations, illustrating how adaptive selection works. The aggregate paired results show the average gains, and the category-level cases identify the templates associated with those gains. Together, they show that VGLW can assign suitable layer weights to different support sets using a single global selection rule.

AuthorsOpenReview ↗

Response to Reviewer Q6Zt (Q3)

Q3.

· Support-set selection

We thank the reviewer for asking about the support sets. In the original manuscript, we use one fixed set of support images per category for each shot setting on each dataset and report point estimates for that configuration. For MVTec-AD, we sort the normal training images in each category by filename and select the first KK images. For VisA, we select KK normal training images using the fixed order in the data split CSV and a preset offset.

· Sharing support sets across ablation configurations

We thank the reviewer for asking about the support sets. In the original manuscript, we use one fixed set of support images per category for each shot setting on each dataset and report point estimates for that configuration. For MVTec-AD, we sort the normal training images in each category by filename and select the first (K) images. For VisA, we select (K) normal training images using the fixed order in the data split CSV and a preset offset. At a given shot count, the ablation configurations share the same support and query sets, with only the module or weighting setting under study changed. The query set also stays fixed across the 1-shot, 2-shot, and 4-shot settings so that the effect of support-set size can be evaluated on the same test samples.

· Uncertainty across repeated evaluations and hyperparameter selection

We thank the reviewer for suggesting that we report statistics across repeated draws. For each evaluation, we first compute the category macro-average for each metric and then report the mean ± standard deviation across three random seeds. Across the two datasets and three shot settings, the standard deviations range from 0.07 to 0.55 percentage points for I-AUROC and from 0.03 to 0.17 percentage points for P-AUROC. The largest observed standard deviation is 0.72 percentage points for 2-shot PRO on MVTec-AD. Increasing the number of support images per category from 1 to 4 improves P-AUROC by 0.49 percentage points on MVTec-AD and 0.48 percentage points on VisA and improves PRO by 0.97 and 1.06 percentage points, respectively. Across the draws examined, pixel-level AUROC varies little, and additional normal support images consistently improve mean region localization performance on both datasets.

During method development, we selected the VGLW thresholds, the DPRS top proportion and ranking smoothing scale, and the PSSA smoothing scale, window size, and mixing weight together based on evaluations on independent development or validation data. We held these settings fixed across both datasets, all categories, and all shot settings. Test performance on MVTec-AD and VisA did not inform their selection.

MVTec-AD

SettingI-AUROCI-AUPRP-AUROCPRO
1-shot97.69 ± 0.0798.98 ± 0.0997.61 ± 0.0694.77 ± 0.30
2-shot97.40 ± 0.1898.89 ± 0.1697.88 ± 0.0395.23 ± 0.72
4-shot98.31 ± 0.4399.439 ± 0.3198.10 ± 0.1795.74 ± 0.16

VisA

SettingI-AUROCI-AUPRP-AUROCPRO
1-shot94.63 ± 0.3095.14 ± 0.1698.18 ± 0.1393.59 ± 0.43
2-shot95.13 ± 0.5595.44 ± 0.1798.48 ± 0.0794.23 ± 0.27
4-shot95.68 ± 0.4695.32 ± 0.4398.66 ± 0.0394.65 ± 0.23
AuthorsOpenReview ↗

Response to Reviewer Q6Zt (Q4)

Q4.

· Resolution used to compute pixel-level metrics

We thank the reviewer for asking about the model input resolution and the pixel-level evaluation resolution. The model input resolution is 672 × 672. We apply spatial processing to the anomaly response maps at that resolution and then resize them to 224 × 224 by bilinear interpolation. We compute all pixel-level metrics on this common evaluation grid. For P-AUROC, we pool pixels from all test images within each category and then take the macro-average of the category-level results. The revised manuscript will specify both resolutions and the resizing step so that readers can reproduce the full procedure from model input to pixel-level evaluation.

· Resizing and binarization of ground-truth masks

We thank the reviewer for asking how the ground-truth masks are processed. On both datasets, the masks are resized to 224 × 224 to align with the anomaly response maps used for evaluation. For MVTec-AD, we explicitly use nearest-neighbor interpolation and binarize the masks at a threshold of 0.5. The original evaluation script for VisA uses Pillow’s default resizing method and then marks nonzero pixels as anomalous. We will describe both procedures separately in the evaluation protocol so that readers can verify the annotation grid and binarization rule used for the pixel-level metrics.

· PRO integration and aggregation across categories

We thank the reviewer for asking us to clarify how PRO is computed. For each category and threshold, we compute the fraction of pixels predicted as anomalous within each ground-truth connected component and then take an equally weighted average over components. This aggregation evaluates individual anomaly regions and prevents larger components from receiving greater weight solely because they contain more pixels. We compute the false positive rate (FPR) using ground-truth background pixels from all test images in the category.

The original implementation scans 1,000 evenly spaced thresholds and retains the samples with FPR ≤ 0.3. It normalizes the retained FPR values by their maximum and integrates the resulting curve using the trapezoidal rule. Finally, it takes the macro-average of the per-category integrals. We will add the threshold sweep, FPR normalization, and two-level aggregation procedure to the evaluation protocol to make the results reproducible.

· Discrepancy in the mean 1-shot P-AUROC in Table A5

We thank the reviewer for pointing out the discrepancy between the category rows and the mean row in Table A5. We checked the column against the original evaluation logs and confirmed that the 1-shot P-AUROC should be 96.1% for metal_nut and 92.8% for transistor. Using these two values and giving all 15 categories equal weight at the precision shown in the table yields approximately 97.63%. Rounded to one decimal place, the result is 97.6%, which matches the reported mean. We will correct the two category rows in the revised manuscript so that they agree with the aggregate result.

Category1-shot I-AUROC1-shot P-AUROC2-shot I-AUROC2-shot P-AUROC4-shot I-AUROC4-shot P-AUROC
bottle100.098.9100.098.9100.098.9
cable95.897.097.097.696.897.7
capsule94.198.095.098.196.698.3
carpet100.099.2100.099.3100.099.3
grid100.099.2100.099.3100.099.3
hazelnut97.799.398.999.499.599.4
leather100.099.2100.099.2100.099.2
metal_nut99.996.1100.096.5100.097.2
pill95.694.896.695.096.195.2
screw88.199.390.699.492.099.4
tile100.097.3100.097.3100.097.4
toothbrush98.698.898.998.998.198.8
transistor98.592.899.394.899.695.2
wood99.296.899.496.899.496.8
zipper99.897.799.897.999.897.9
Mean97.897.698.397.998.598.0
Reviewer_WsQkOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes HIVE-AD, a training-free framework for few-shot industrial anomaly detection based on frozen DINOv2 features. The main idea is to exploit information from multiple backbone depths rather than relying on a single representation or a fixed aggregation rule. The framework has three main components. VGLW estimates the spatial variability of normal support features and selects one of three predefined layer-weight profiles. DPRS separates the evidence used for image-level anomaly ranking from that used for pixel-level localization. PSSA further smooths and spatially expands the localization response. The whole visual pipeline is deterministic and requires no parameter optimization for a new category. An optional frozen MLLM converts localized regions into coarse structured defect reports.

The method is evaluated on MVTec-AD and VisA under 1-shot, 2-shot, and 4-shot settings. In the 1-shot case, HIVE-AD reaches 97.8 image AUROC and 97.6 pixel AUROC on MVTec-AD, and 94.5 and 98.2 respectively on VisA. The results are competitive with or better than recent few-shot methods, although, as a minor observation, SubspaceAD remains stronger on VisA PRO.

Critical strengths

S1. Simple and strong as training free. The method does not optimize prompts, adapters, feature memories, or category-specific parameters. This is an attractive property for few-shot industrial inspection. The gains over the authors' own frozen DINOv2-G baseline are also meaningful: image-level AUROC improves from 95.5 to 97.8 on MVTec-AD and from 91.7 to 94.5 on VisA. Importantly, the ablations indicate that the improvement does not come from a single isolated trick. VGLW, DPRS, and PSSA affect different aspects of the result, and their combinations generally improve performance further. The numbers speak quite clearly here, and I find this part convincing.

S2. Decoupling image-level ranking from localization is simple but well-motivated. Actually, image classification and dense localization have different requirements. A few strong responses can be sufficient for ranking an image as anomalous, while localization requires a spatially coherent map. Using separate readouts for these two tasks is sensible, and the results support this choice. DPRS gives a particularly clear improvement on VisA image-level detection, while PSSA improves region-based localization metrics.

The qualitative maps are also generally clean. The examples on pages 8 and 20 to 24 show compact anomaly responses and relatively limited background activation across quite different object and texture categories.

Critical weaknesses

W1. The novelty of the core method is somewhat limited, and the most distinctive component, VGLW, remains heuristic. But this is not a very big problem to me, due to the results. At its core, HIVE-AD performs nearest-neighbor anomaly scoring on frozen multilevel DINOv2 features, followed by weighted aggregation, smoothing, top-ratio pooling, and local max filtering. These are individually simple operations. The main conceptual novelty therefore rests heavily on how these operations are organized. 1) VGLW is the component that is supposed to adapt representation depth to the support set: however, the mapping from spatial variance to depth is not derived from an analysis of which layers are actually optimal. 2) The method computes one scalar from the deepest selected layer and then chooses among three manually specified weight templates using two fixed thresholds. For me, this is where the game is played. The sensitivity study shows that the method is not very sensitive to these choices, which is good. But robustness to the thresholds is different from demonstrating that the proposed statistic is actually predictive of which feature depths should be emphasized. The authors themselves acknowledge that the variance-to-depth mapping remains heuristic... and I appreciate it. Anyway,I think the paper needs a more direct experiment here. For example, one could compare the VGLW-selected profile with the oracle-best layer profile for each category or support set and measure whether the spatial-variance statistic genuinely predicts the optimal representation depth.

W2. The headline comparison with previous methods is not sufficiently controlled to cleanly establish the advantage of HIVE-AD. Table 1 compares complete published systems using their original backbones, input resolutions, and post-processing. This is useful for practical positioning, but it mixes improvements due to the HIVE-AD design with differences in backbone capacity and evaluation configuration. This issue is especially relevant because HIVE-AD uses DINOv2-G at 672 × 672 resolution, while performance improves considerably with backbone scale. The paper itself shows that larger DINOv2 variants generally improve image-level ranking. The comparison with SubspaceAD is therefore difficult to interpret mechanistically. HIVE-AD is better on image-level metrics, while SubspaceAD remains better on VisA PRO, but it is unclear how much of this difference would remain if the two approaches used exactly the same DINOv2 backbone, feature layers, resolution, support samples, and spatial post-processing. For an ICLR paper whose main contribution is how frozen DINOv2 evidence is organized, I think a controlled comparison against the strongest DINOv2-based competitors would substantially strengthen the empirical argument.

Questions for authors

Q1. Does the support spatial-variance score actually predict which DINOv2 layers are most useful for a category? Could the authors compare the VGLW-selected profile with an oracle profile obtained by evaluating the alternative layer weightings separately? I think this would provide much stronger evidence for the central VGLW hypothesis than threshold sensitivity alone.

Q2. Can the authors provide a controlled comparison with AnomalyDINO and, especially, SubspaceAD using the same DINOv2 backbone, image resolution, support samples, and ideally the same selected layers? This would help separate the effect of HIVE-AD's evidence organization from the effect of backbone scale and implementation choices.

Q3. How were the global hyperparameters, in particular the VGLW thresholds, the top-ratio (r), the two Gaussian smoothing scales, and the PSSA window size, selected? The fact that one configuration is used for all datasets and categories is a strength, but it would be important to know whether MVTec-AD or VisA test performance influenced this configuration at any point.

Q4. The optional MLLM branch has only moderate automatic report accuracy, around 60% to 68% depending on the attribute, and the correctness of the free-form rationale is not evaluated. Would the authors consider reducing the emphasis on this component, or alternatively validating it with a blinded human evaluation? At present I see it mainly as an interface demonstration rather than a validated contribution.

19 comments
AuthorsOpenReview ↗

Response to Reviewer WsQk (Q1)

We thank the reviewer for the encouraging assessment, for recognizing our design and experimental results in training-free few-shot industrial anomaly detection and the decoupling of image ranking from pixel localization, and for raising questions about layer-weight selection, controlled comparisons, global hyperparameter selection, and evaluation of the optional reporting branch.

Q1.

· Comparison of VGLW-selected layer weights with the per-draw oracle

We thank the reviewer for suggesting an oracle comparison to directly test layer-weight selection. For the same category and the same support-set draw, we evaluate the early, uniform, and late templates separately, keeping all other detection settings fixed. The candidate configuration with the highest image-level AUROC serves as the oracle for that draw. If the configuration selected by VGLW ties with another candidate for the highest AUROC, it also counts as a hit. During detection, VGLW determines layer weights solely from the spatial variance of the normal support set. As shown in the table below, VGLW achieves image-level AUROCs of 96.05% on MVTec-AD and 93.29% on VisA, with gaps of 0.33 and 0.53 percentage points to the per-draw oracle, respectively. This comparison connects the basis for layer-weight selection to the performance of the selected configuration, showing that normal support features alone can yield detection results close to the per-draw optimum among the three candidate templates.

MethodMVTec I-AUROCVisA I-AUROC
VGLW96.05 ± 1.0993.29 ± 0.96
Best oracle per category and draw96.38 ± 0.9993.82 ± 0.96

· Whether support spatial variance can identify suitable weight configurations for specific categories

We thank the reviewer for suggesting that we examine how support variance relates to the gains from layer weighting at the category level. The category-level case analysis uses 1-shot support sets generated with three random seeds. Across the three support sets, VGLW consistently selects uniform for wood and late for screw, toothbrush, pill, and zipper. Each of these five categories achieves the highest image-level AUROC among the candidate templates in every draw. For screw, the selected late template improves mean image-level AUROC by 9.34 percentage points over uniform. On VisA, both chewinggum and macaroni2 match the per-draw oracle in two of the three draws. The mean gap is 0.02 percentage points for chewinggum and 1.07 percentage points for macaroni2. Categories such as wood and screw achieve their respective best results with different templates, illustrating how support-guided fusion can accommodate the feature-depth preferences of some product categories.

CategoryVGLW SelectionVGLW I-AUROCPer-Draw Oracle I-AUROCOracle Hits
MVTec/wooduniform99.42 ± 0.1899.42 ± 0.183/3
MVTec/screwlate79.03 ± 3.1779.03 ± 3.173/3
MVTec/toothbrushlate99.72 ± 0.4899.72 ± 0.483/3
MVTec/pilllate96.48 ± 1.1596.48 ± 1.153/3
MVTec/zipperlate99.93 ± 0.0599.93 ± 0.053/3
VisA/chewinggumlate98.75 ± 0.8298.77 ± 0.842/3
VisA/macaroni2late83.58 ± 0.7284.65 ± 2.032/3
AuthorsOpenReview ↗

Response to Reviewer WsQk (Q2)

Q2.

We thank the reviewer for suggesting a comparison of HIVE-AD, AnomalyDINO, and SubspaceAD under matched feature extraction. We added a comparative experiment in which all three methods use the same frozen DINOv2-G backbone, a 672 × 672 input resolution, and the same five feature blocks [11, 17, 23, 29, 39]. For the 1-shot, 2-shot, and 4-shot settings on both datasets, we evaluated each method with three random seeds and report the mean ± standard deviation. Under this protocol, the methods share feature extraction conditions and normal reference samples, so the comparison more directly reflects how each method fuses and scores feature evidence. Reporting repeated evaluations separately for each shot setting also makes it possible to examine whether the relative performance of the methods changes with the number of support images.

MVTec-AD

SettingMethodI-AUROCI-AUPRP-AUROCPRO
1-shotHIVE-AD (ours)97.69 ± 0.0798.98 ± 0.0997.61 ± 0.0694.77 ± 0.30
1-shotAnomalyDINO96.40 ± 0.1598.26 ± 0.1396.95 ± 0.0894.32 ± 0.27
1-shotSubspaceAD94.89 ± 0.1797.01 ± 0.1597.41 ± 0.1394.09 ± 0.19
2-shotHIVE-AD (ours)97.40 ± 0.1898.89 ± 0.1697.88 ± 0.0395.23 ± 0.72
2-shotAnomalyDINO96.79 ± 0.2098.48 ± 0.2597.31 ± 0.0394.97 ± 0.87
2-shotSubspaceAD95.78 ± 0.3997.67 ± 0.4297.77 ± 0.0194.69 ± 0.32
4-shotHIVE-AD (ours)98.31 ± 0.4399.439 ± 0.3198.10 ± 0.1795.74 ± 0.16
4-shotAnomalyDINO96.60 ± 0.4998.35 ± 0.4996.59 ± 0.1092.61 ± 0.88
4-shotSubspaceAD97.07 ± 0.9198.30 ± 0.8697.01 ± 0.1194.19 ± 0.28

VisA

SettingMethodI-AUROCI-AUPRP-AUROCPRO
1-shotHIVE-AD (ours)94.63 ± 0.3095.14 ± 0.1698.18 ± 0.1393.59 ± 0.43
1-shotAnomalyDINO94.52 ± 0.3295.05 ± 0.0497.27 ± 0.2393.22 ± 0.33
1-shotSubspaceAD94.40 ± 0.5494.30 ± 0.5698.00 ± 0.0195.50 ± 0.15
2-shotHIVE-AD (ours)95.13 ± 0.5595.44 ± 0.1798.48 ± 0.0794.23 ± 0.27
2-shotAnomalyDINO94.58 ± 0.5793.89 ± 0.4997.65 ± 0.1393.70 ± 0.23
2-shotSubspaceAD93.87 ± 0.4294.56 ± 0.3697.80 ± 0.3395.63 ± 0.19
4-shotHIVE-AD (ours)95.68 ± 0.4695.32 ± 0.4398.66 ± 0.0394.65 ± 0.23
4-shotAnomalyDINO91.97 ± 0.5592.07 ± 0.5797.90 ± 0.0493.19 ± 0.53
4-shotSubspaceAD94.31 ± 0.4794.09 ± 0.7197.93 ± 0.0395.31 ± 0.22

On MVTec-AD, HIVE-AD achieves the highest mean on all four metrics in each of the three shot settings. For example, in the 1-shot setting, its I-AUROC is 97.69%, exceeding AnomalyDINO and SubspaceAD by 1.29 and 2.80 percentage points, respectively. It also achieves a P-AUROC of 97.61% and a PRO of 94.77%. These results show that the gains in both image ranking and pixel localization persist with a shared backbone and support set. On VisA, HIVE-AD achieves the highest mean I-AUROC, I-AUPR, and P-AUROC in all three shot settings. Its 4-shot I-AUROC and P-AUROC reach 95.68% and 98.66%, respectively. Its Pro is close to the results of the other 2 methods. Both tables include PRO so that image ranking, pixel ranking, and region coverage can be examined separately. Overall, the results show that HIVE-AD’s organization of multilevel evidence supports both image-level detection and pixel-level anomaly localization using the same frozen features.

AuthorsOpenReview ↗

Response to Reviewer WsQk (Q3)

Q3.

· Selection of global hyperparameters

We thank the reviewer for asking how the global hyperparameters were selected. During method development, we determined a single configuration based on evaluations on independent development or validation data, considering the role of each parameter in the detection pipeline. We set the VGLW thresholds to 0.8 and 1.5 to assign weights to features from different depths according to the spatial variance of normal support features. The image-ranking branch uses a Gaussian smoothing scale of σᵣ = 4.0 to suppress isolated responses and a top proportion of r = 0.002 to emphasize high anomaly responses in the ranking score. The localization branch uses a Gaussian smoothing scale of σₚ = 8.0, a PSSA window size of h = 31, and a mixing weight of α = 0.6 to control smoothing and spatial expansion of the dense response map. The final configuration remains fixed across all categories and shot settings on MVTec-AD and VisA. The fixed thresholds provide a common selection rule, and the resulting layer weights depend on the normal support features. This single global configuration therefore adapts feature fusion to the support set without requiring a new search for these hyperparameters for each new product category.

· Use of test performance in hyperparameter selection

We thank the reviewer for asking about the relationship between hyperparameter selection and test evaluation. We selected all hyperparameters using independent development or validation data, with no use of test performance on MVTec-AD or VisA. We then fixed the configuration for test evaluation. Table 3 reports sensitivity to different VGLW thresholds and to a continuous mapping variant. Across the settings in the table, I-AUROC varies by 0.45 percentage points on MVTec-AD and 0.54 percentage points on VisA, indicating limited variation in image-ranking performance over the configurations examined.

MVTec-AD

Mapping(τ₁, τ₂)I-AUROCP-AUROCPRO
Discrete(0.6, 1.8)97.3897.3494.12
Discrete†(0.8, 1.5)97.8297.6494.61
Discrete(1.0, 1.3)97.8297.6494.61
Continuous(0.6, 1.8)97.8197.6194.59
Continuous(0.8, 1.5)97.8397.6294.55
Continuous(1.0, 1.3)97.8197.6094.50

VisA

Mapping(τ₁, τ₂)I-AUROCP-AUROCPRO
Discrete(0.6, 1.8)94.1197.9093.06
Discrete†(0.8, 1.5)94.5498.1793.38
Discrete(1.0, 1.3)94.2798.0593.23
Continuous(0.6, 1.8)94.6598.0393.13
Continuous(0.8, 1.5)94.2998.0693.24
Continuous(1.0, 1.3)94.2798.0593.23

† denotes the default configuration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.