Pith. sign in

REVIEW 4 major objections 4 minor 43 references

FastRef:Fast Prototype Refinement for Few-Shot Industrial Anomaly Detection

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read FastRef is a test-time procedure that refines a model's normal prototypes using the query image itself, and reports consistent few-shot anomaly detection gains on four benchmarks.

desk verdict Plausible test-time prototype refinement for few-shot IAD, but the central empirical claim rests on single-run AUROC gains that are often under one point; needs repeated-seed evaluation and a clearer boundary with FastRecon and FOCT. read the letter →

arxiv 2506.21398 v1 pith:F3ABU6WB submitted 2025-06-26 cs.CV

classification cs.CV
keywords few-shotanomalydetectionindustrialprototyperefinementoptimaltransportSinkhornalgorithmtest-timeadaptationlocalizationfeaturereconstruction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

FastRef is a test-time procedure for few-shot industrial anomaly detection that refines the small set of normal prototypes on the fly, using the features of the very image being inspected. The paper claims that alternating between transferring query characteristics into the prototypes and suppressing anomalies with an optimal-transport alignment step consistently improves both image-level and pixel-level detection. The gains are demonstrated by plugging FastRef into PatchCore, FastRecon, WinCLIP, and AnomalyDINO across MVTec, VisA, MPDD, and RealIAD under 1, 2, and 4 normal support images. A sympathetic reader should take away that query-image statistics can be used safely at test time, and that the refinement costs only a few milliseconds per image, so the approach is practical for automated inspection settings with rapid product changeovers.

What carries the argument

The load-bearing object is the nested optimization in Eq. 3, along with its alternating solver. The transform matrix $W$ implements what the paper calls composition refinement: each query feature is reconstructed by selecting and combining entries from the prototype bank $M_s$, so the refined prototypes $M_s^w = W M_s$ carry the query's normal statistics into the bank. The transport plan $T$, computed by the Sinkhorn algorithm for entropy-regularized optimal transport, is the anomaly-suppression device: it measures the distributional gap between refined and original prototypes, and because the original prototypes are normal, minimizing that gap cancels any anomalous components the reconstruction step would otherwise copy. The claim that these two roles balance each other is what the experiments test; the paper's convergence argument for the alternating scheme relies on the $W$-subproblem being strongly convex.

What would settle it

Construct a set of query images whose defects are built from the same local patches as the normal support images, such as a normal texture patch that is misplaced or overlaid, and run FastRef; if the refined prototypes shift toward a defective query and the anomaly score on a subsequent truly normal query drops below the unrefined baseline, the transport term has failed to suppress the anomaly. A sharper check is to feed a fully anomalous query and measure how far the refined prototype bank moves from the original: under the paper's claim the movement should be small, and a large movement would falsify the suppression mechanism.

Watch

Extended reading notes

Core claim

The central claim is that in few-shot industrial anomaly detection the normal prototypes built from a handful of support images are the main bottleneck, and that refining them with query statistics, while explicitly suppressing the anomalies those query statistics may contain, is enough to improve detection without retraining. FastRef writes the refinement as a nested objective: minimize $\mathrm{dis}(f_q, W M_s) + \lambda\,\mathrm{OT}(p,q)$, where $W$ is a transformation matrix that reconstructs the query features by composing items from the prototype bank and $T$ is a transport plan that aligns the refined prototypes back to the original normal prototypes. The optimal-transport term is the paper's mechanism for anomaly suppression: because the original prototypes come only from normal images, pulling the refined prototypes toward them prevents anomalous query regions from being copied into the bank. With Euclidean or cosine distance the $W$ update has a closed form, and $T$ can be computed with the Sinkhorn algorithm; the paper reports that two outer iterations and fewer than ten Sinkhorn loops suffice. Across PatchCore, FastRecon, WinCLIP, and AnomalyDINO, the refined versions produced by FastRef consistently beat the unrefined baselines on MVTec, VisA, MPDD, and RealIAD under 1, 2, and 4 shots, with inference times within a few milliseconds of the originals.

Load-bearing premise

The load-bearing premise is that features from a frozen pretrained backbone are arranged so that a query's normal content can be reconstructed from a few normal prototypes while its anomalous content can be identified by the optimal-transport pull-back; if anomalous patches are as easy to reconstruct from normal prototypes as normal patches are, refinement will contaminate the bank and the method's claimed safety collapses.

Editorial extensions

If this is right

  • As a plug-in, FastRef improves PatchCore, FastRecon, WinCLIP, and AnomalyDINO without retraining the feature extractor, so any future prototype-based few-shot detector can adopt the same refinement.
  • The largest reported gains occur where few-shot data is hardest: WinCLIP+ gains 7.0 image-level AUROC points on MPDD at 4 shots and PatchCore+ gains 7.3 points on VisA at 1 shot.
  • Optimal-transport anomaly suppression outperforms a Gaussian test-time-alignment baseline on MPDD at 1, 2, and 4 shots, supporting the paper's claim that the non-Gaussian assumption matters.
  • The method stays real-time: per-image inference on MVTec rises from 0.098 s to 0.100 s for PatchCore and from 0.180 s to 0.184 s for WinCLIP.
  • Refinement helps across CNN, CLIP, and DINOv2 features and across WRN-50, ViT, and EfficientNet backbones, so the gains are not tied to one feature family.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same nested reconstruction-plus-OT step could be applied to any memory-based detector whose bank of normal features is fixed, including full-shot methods, as a cheap test-time adaptation to distribution shift; the paper does not test that setting.
  • Because $\lambda$ and the coreset ratio are tuned per dataset, one testable extension is a principled rule for setting them from support-set size or backbone feature dimensionality; the paper's sensitivity curves show a peak rather than a plateau, so such a rule could matter in practice.
  • The convergence proof assumes the $W$-subproblem is strongly convex, which requires $M_s M_s^T$ to be positive definite; with very few support images the prototype bank may be rank-deficient, so monitoring the condition number of the inverse in the update rule is a robustness check the paper does not perform.
  • One stronger distribution shift to test is cross-class refinement, where support and query come from different product classes; the ablation suggests $W$ alone can inject anomalies, so it is unclear how much OT suppression would help there.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes FastRef, a test-time prototype refinement framework for few-shot industrial anomaly detection. FastRef iterates between two steps: transferring query-image characteristics into the normal prototypes through a learned linear transformation matrix W, and suppressing anomalies by aligning the refined prototypes to the original normal prototypes with entropy-regularized optimal transport. The method is designed as a plug-in for prototype-oriented FS-IAD methods, and the authors instantiate it as PatchCore+, FastRecon+, WinCLIP+, and AnomalyDINO+. Experiments on MVTec, VisA, MPDD, and RealIAD under 1/2/4 shots report image- and pixel-level AUROC improvements, an ablation study, hyper-parameter analyses, and an inference-time comparison, together with a convergence proof in the appendix. The paper claims that the refined prototypes lead to consistent performance gains over the baselines while maintaining real-time efficiency.

Significance. If the empirical claims are sound, FastRef offers a simple and fast test-time adaptation strategy that could benefit a range of prototype-based few-shot anomaly detection systems, and the idea of using optimal transport to suppress anomaly contamination during refinement is interesting and potentially useful beyond the specific methods tested. The paper includes a closed-form update derived in an appendix, a convergence proof, and a comparison against a Gaussian-assumption alternative, which are positive features. However, the strength of the contribution is currently limited by the fragility of the empirical evidence: all reported AUROC values are single runs with no variance across support-set draws, several reported gains are under one point, one of the four claimed integrations (AnomalyDINO+) is never evaluated, and the theoretical derivation for the cosine-distance variant has a gap. The central claim of 'consistent superiority' is therefore not yet established; the work would be much more convincing with repeated-seed evaluation, a corrected or qualified derivation, and an honest scope of experiments.

major comments (4)
  1. [§6.2, Tables 1-2] The central empirical claim that PatchCore+/WinCLIP+ 'consistently delivers superior IAD performance' rests entirely on single-run AUROC values. Many decisive deltas are very small (e.g., WinCLIP+ gains on MVTec image-level are +0.2, +0.2, and +0.2 for 1/2/4 shots; several pixel-level gains are also below one point). In few-shot IAD, random selection of the k support images typically induces AUROC variation of several points, so a single run cannot distinguish a real improvement from sampling noise. Please report means and standard deviations (or confidence intervals) over at least 5-10 random support-set draws, and state the number of draws used.
  2. [§5.4 and §6] AnomalyDINO+ is introduced in Section 5.4 as one of the integrated methods, and the abstract/introduction say FastRef is applied to 'PatchCore, FastRecon, WinCLIP, and AnomalyDINO'. However, no experiment with AnomalyDINO+ appears anywhere in Section 6, Tables 1-2, or the ablation studies. Either add the missing experiments or explicitly remove AnomalyDINO+ from the claims of integration and evaluation.
  3. [Appendix A and Eq. (5)] The closed-form update in Eq. (5) is derived for the Euclidean-distance objective in Appendix A. For WinCLIP+ (Section 5.3) the distance in the reconstruction term and presumably the OT cost are cosine distances, for which the objective is not quadratic and the Euclidean minimizer is not the minimizer of the cosine objective once the OT term is included. The appendix's argument that 'both Euclidean distance and cosine distance achieve minimum when f_q(i,:) = W_i,: M_s' considers only the reconstruction term and ignores the λ OT(p,q) term. This makes Eq. (5) inapplicable to the WinCLIP+ variant as stated; please provide a correct derivation or explicitly justify Eq. (5) as an approximation for the cosine case.
  4. [Appendix C] The convergence proof relies on F(W) being µ-strongly convex, which requires M_s M_s^T to be positive definite. With Coreset-sampled prototypes, M_s is n×c; when n > c (which is typical after dense feature extraction), M_s M_s^T is rank-deficient and strong convexity does not hold. The proof does not state or justify assumptions on the prototype matrix, and the claim that the algorithm converges for the actual settings used in the experiments is therefore unsupported. Please state the conditions under which the proof is valid, or provide a convergence argument that does not depend on the positive-definiteness of M_s M_s^T.
minor comments (4)
  1. [Abstract and §5] The abstract and Section 5 say FastRef is integrated with 'three ... methods' but then list four names (PatchCore, FastRecon, WinCLIP, AnomalyDINO); please correct this inconsistency.
  2. [Eq. (5)] The denominator '1 + λT_l·1' is ambiguous; please specify that the division is elementwise and that T_l·1 is an n-dimensional vector, or rewrite the equation with explicit row-wise operations.
  3. [§6.1 and Fig. 6] Hyper-parameter analyses for α, λ, and L are shown only for MVTec under 4 shots; a statement about whether the selected values transfer to other datasets would help, especially since different α values are used per dataset in the implementation details.
  4. [Throughout] Please correct typos and inconsistent spellings: 'ViSA' should be 'VisA' (also in the abstract and Tables), 'CoreSet' vs 'Coreset', 'paradiagm' (intro), and 'competitve' (Section 6.5).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FastRef's refinement objective and anomaly score form a self-contained transductive optimization whose benchmark evaluation is external; the same-query fitting is a design choice rather than an equivalence-by-construction prediction.

full rationale

The central derivation, Eq. 3, defines W* and T* by jointly minimizing reconstruction of the query features f_q from refined prototypes W M_s and an optimal-transport penalty that ties W M_s to the original normal prototypes M_s. The anomaly score in Eq. 6 is the residual of that reconstruction, so the method is explicitly transductive: the query is used to refine the prototypes that later score the same query. That is a legitimate test-time adaptation design, not a circular derivation, because the OT term is anchored to normal prototypes rather than to the query's own label, and because the reported AUROC values are computed against external anomaly labels on MVTec, VisA, MPDD, and RealIAD. No parameter is fitted to those labels, and the closed-form update in Eq. 5 is derived from Eq. 3 rather than imported from a self-citation. The paper's self-citation to FOCT (Ref. [34]) appears only as a baseline in the experiments and related work and is not load-bearing. Several correctness concerns exist but are not circularity: the claim that cosine and Euclidean distances share the same minimizer (Appendix A) is mathematically unjustified, the convergence proof (Appendix C) assumes strong convexity of F(W) without guaranteeing positive definiteness of M_s M_s^T, the empirical gains are reported without repeated-seed variance, and AnomalyDINO+ is not experimentally evaluated. These are substantive scientific concerns, but none of them makes the derivation equivalent to its own inputs by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim relies on the pre-trained feature space separating normal from anomalous content, on transductive query/support distribution matching, and on a strong convexity condition in the convergence proof. Free parameters are the regularization weight lambda, the Coreset ratio alpha, the number of outer iterations L, Sinkhorn inner iterations, and the entropy coefficient epsilon.

free parameters (5)
  • lambda = 0.3 for PatchCore/FastRecon, 0.1 for WinCLIP
    Balance coefficient between reconstruction and OT terms in Eq. 3, tuned per method.
  • alpha (Coreset ratio) = 0.05 for PatchCore+, 0.5/0.3/0.2/0.1 for WinCLIP+ on MVTec/VisA/MPDD/RealIAD
    Controls prototype size and representativeness, tuned per dataset for WinCLIP+.
  • L (outer iterations) = 2
    Number of alternating W/T updates, selected for efficiency-performance trade-off.
  • Sinkhorn inner iterations = about 10
    Inner loops for transport probability update; exact count unspecified.
  • epsilon (entropy regularization) = not reported
    Entropy coefficient in Eq. 4 for Sinkhorn; value never given.
assumptions (4)
  • domain assumption Pre-trained feature extractors (WRN-50, CLIP ViT-B/16+, DINOv2 small) provide features where normal and anomalous content are distinguishable.
    The entire method operates in the feature space of these backbones (Sec 5, Sec 6.1) and the claim of improved detection rests on this.
  • domain assumption In the few-shot regime, normal prototypes from k shots form a sufficient linear basis for reconstructing query features while anomalies can be suppressed by the OT prior.
    Eq. 3 assumes W*Ms can represent query features, and the OT term assumes closeness to the original prototypes prevents anomaly copying.
  • domain assumption Query and support images share the same feature distribution (transductive assumption).
    Test-time refinement in Sec 4.1 uses query statistics to adjust prototypes, which only helps if query features are informative about the normal manifold.
  • standard math F(W) is mu-strongly convex, requiring M_s M_s^T to be positive definite.
    Appendix C invokes strong convexity from [2] without proving that the Coreset-sampled prototype matrix M_s has full row rank.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FastRef:Fast Prototype Refinement for Few-Shot Industrial Anomaly Detection." pith.science (2026). https://pith.science/paper/F3ABU6WB

@misc{pith2026250621398,
  author       = {Pith},
  title        = {Pith review of: FastRef:Fast Prototype Refinement for Few-Shot Industrial Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F3ABU6WB}},
  note         = {Machine review of arXiv:2506.21398}
}
read the original abstract

Few-shot industrial anomaly detection (FS-IAD) presents a critical challenge for practical automated inspection systems operating in data-scarce environments. While existing approaches predominantly focus on deriving prototypes from limited normal samples, they typically neglect to systematically incorporate query image statistics to enhance prototype representativeness. To address this issue, we propose FastRef, a novel and efficient prototype refinement framework for FS-IAD. Our method operates through an iterative two-stage process: (1) characteristic transfer from query features to prototypes via an optimizable transformation matrix, and (2) anomaly suppression through prototype alignment. The characteristic transfer is achieved through linear reconstruction of query features from prototypes, while the anomaly suppression addresses a key observation in FS-IAD that unlike conventional IAD with abundant normal prototypes, the limited-sample setting makes anomaly reconstruction more probable. Therefore, we employ optimal transport (OT) for non-Gaussian sampled features to measure and minimize the gap between prototypes and their refined counterparts for anomaly suppression. For comprehensive evaluation, we integrate FastRef with three competitive prototype-based FS-IAD methods: PatchCore, FastRecon, WinCLIP, and AnomalyDINO. Extensive experiments across four benchmark datasets of MVTec, ViSA, MPDD and RealIAD demonstrate both the effectiveness and computational efficiency of our approach under 1/2/4-shots.

Figures

Figures reproduced from arXiv: 2506.21398 by the authors.

Figure 1
Figure 1. Left: The pipeline of traditional prototype-oriented few-shot IAD methods. Right: Our proposed framework of FastRef, it refines prototypes with the current query features by iterating between suppressing anomalies and transferring characteristics. The improved anomaly detection performance could be obtained thanks to the well refined prototypes. measurement functions. Anomalies are detected by comparing the anomaly … view at source ↗
Figure 2
Figure 2. The roles of transformation matrix W and transport probabil￾ity T . Top: W-based f q t reconstruction restores normal features but may propagate anomalies, causing similar anomaly score (AScore) in defective and normal regions. Bottom: T -adjusted W enables effective anomaly suppres￾sion, making defective regions’ AScore distinguishable. The indexes between GT mask and T show that T primarily activates on anomalous … view at source ↗
Figure 3
Figure 3. From left to right are results of pixel-level anomaly localization with MVTec, ViSA, MPDD, and RealIAD under 4-shots. 5. Application to Prototype-oriented Methods we apply the proposed FastRef model in Sec. 4 to three popular and recently developed prototype-oriented few-shot IAD methods, including PatchCore [27], FastRecon [7], WinCLIP [13], and AnomalyDINO [5]. Next, we will discuss them one by one. 5.1. PatchCore… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: From top to bottom are results on ablation studies of 2-shots anomaly localization on MVTec and VisA datasets. AUROC on the MPDD dataset under 4-shots. Furthermore, the improvement delivered by our model surpasses that of the point-to-point regularization approach used…
Figure 5
Figure 5. Figure 5: More visualization results of pixel-level anomaly localization on MVTec, ViSA, MPDD, and RealIAD under 4-shots [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: From left to right are analysis on hyper-parameters of α, λ, and iteration number with MVTec dataset under 4-shots [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: The transform matrix in (b) cor￾responding to the bottle in (a). (c) shows the 2-D PCA of Top-3 selected items in the normal prototypes according to (b) for fore￾ground, background, and anomaly regions of query image in (a) [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 29 canonical work pages

  1. [1]

    Mvtec ad–a com- prehensive real-world dataset for unsupervised anomaly detection

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad–a com- prehensive real-world dataset for unsupervised anomaly detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592–9600, 2019

  2. [2]

    Convex optimization: Algorithms and complexity.Foundations and Trends®in Machine Learning, 8(3-4):231–357, 2015

    Sébastien Bubeck et al. Convex optimization: Algorithms and complexity.Foundations and Trends®in Machine Learning, 8(3-4):231–357, 2015

  3. [3]

    Sub-image anomaly detection with deep pyramid corre- spondences.arXiv preprint arXiv:2005.02357, 2020

    Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid corre- spondences.arXiv preprint arXiv:2005.02357, 2020

  4. [4]

    Sinkhorn distances: Lightspeed computation of optimal transport.Ad- vances in neural information processing systems, 26, 2013

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport.Ad- vances in neural information processing systems, 26, 2013

  5. [5]

    Anomalydino: Boostingpatch-basedfew-shotanomalydetectionwithdinov2

    Simon Damm, Mike Laszkiewicz, Johannes Lederer, and Asja Fischer. Anomalydino: Boostingpatch-basedfew-shotanomalydetectionwithdinov2. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1319–1329. IEEE, 2025

  6. [6]

    Padim: a patch distribution modeling framework for anomaly detection and localization

    Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. InInter- national Conference on Pattern Recognition, pages 475–489. Springer, 2021

  7. [7]

    Fas- trecon: Few-shot industrial anomaly detection via fast feature reconstruction

    Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qiugui Hu, and Jimin Xiao. Fas- trecon: Few-shot industrial anomaly detection via fast feature reconstruction. InProceed- ings of the IEEE/CVF International Conference on Computer Vision, pages 17481–17490, 2023

  8. [8]

    Memorizingnormalitytodetectanomaly: Memory- augmented deep autoencoder for unsupervised anomaly detection

    Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, andAntonvandenHengel. Memorizingnormalitytodetectanomaly: Memory- augmented deep autoencoder for unsupervised anomaly detection. InProceedings of the IEEE/CVF international conference on computer vision, pages 1705–1714, 2019

Show all 43 references
  1. [9]

    Anomalygpt: Detecting industrial anomalies using large vision-language models

    Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision-language models. InPro- ceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1932–1940, 2024

  2. [10]

    Diad: A diffusion-based framework for multi-class anomaly detection

    Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. Diad: A diffusion-based framework for multi-class anomaly detection. 2023. URLhttps://arxiv.org/abs/2312.06607

  3. [11]

    Maskr-cnn

    KaimingHe, GeorgiaGkioxari, PiotrDollár, andRossGirshick. Maskr-cnn. InProceedings of the IEEE international conference on computer vision, pages 2961–2969, 2017

  4. [12]

    Registration based few-shot anomaly detection

    Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan- Feng Wang. Registration based few-shot anomaly detection. InEuropean Conference on Computer Vision, pages 303–319. Springer, 2022

  5. [13]

    Winclip: Zero-/few-shot anomaly classification and segmentation

    Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. Winclip: Zero-/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606–19616, 2023

  6. [14]

    Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions

    Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In2021 13th International congress on ultra modern telecommunications and control systems and worksh...

  7. [15]

    Anomaly detection for predictive maintenance in industry 4.0-a survey

    Pooja Kamat and Rekha Sugandhi. Anomaly detection for predictive maintenance in industry 4.0-a survey. InE3S web of conferences, volume 170, page 02007. EDP Sciences, 2020. 20

  8. [16]

    Promptad: Learning prompts with only normal samples for few-shot anom- aly detection

    Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. Promptad: Learning prompts with only normal samples for few-shot anom- aly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838–16848, 2024

  9. [17]

    Deep industrial image anomaly detection: A survey.Machine Intelligence Research, 21(1):104–135, 2024

    Jiaqi Liu, Guoyang Xie, Jinbao Wang, Shangnian Li, Chengjie Wang, Feng Zheng, and Yaochu Jin. Deep industrial image anomaly detection: A survey.Machine Intelligence Research, 21(1):104–135, 2024

  10. [18]

    Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34:21808–21820, 2021

    Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34:21808–21820, 2021

  11. [19]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012– 10022, 2021

  12. [20]

    Simplenet: A simple network for image anomaly detection and localization

    Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402–20411, 2023

  13. [21]

    Hierarchical vector quantized transformer for multi-class unsupervised anomaly de- tection.arXiv preprint arXiv:2310.14228, 2023

    Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly de- tection.arXiv preprint arXiv:2310.14228, 2023

  14. [22]

    Hierarchical vector quantized transformer for multi-class unsupervised anomaly de- tection.Advances in Neural Information Processing Systems, 36, 2024

    Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly de- tection.Advances in Neural Information Processing Systems, 36, 2024

  15. [23]

    Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

  16. [24]

    The matrix cookbook.Technical University of Denmark, 7(15):510, 2008

    Kaare Brandt Petersen, Michael Syskind Pedersen, et al. The matrix cookbook.Technical University of Denmark, 7(15):510, 2008

  17. [25]

    Computational optimal transport: With applications to data science.Foundations and Trends®in Machine Learning, 11(5-6):355–607, 2019

    Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science.Foundations and Trends®in Machine Learning, 11(5-6):355–607, 2019

  18. [26]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pag...

  19. [27]

    Towards total recall in industrial anomaly detection

    Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318–14328, 2022

  20. [28]

    Optimizing patchcore for few/many-shot anomaly detection.arXiv preprint arXiv:2307.10792, 2023

    João Santos, Triet Tran, and Oliver Rippel. Optimizing patchcore for few/many-shot anomaly detection.arXiv preprint arXiv:2307.10792, 2023

  21. [29]

    Active learning for convolutional neural networks: A core-set approach.arXiv preprint arXiv:1708.00489, 2017

    Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach.arXiv preprint arXiv:1708.00489, 2017

  22. [30]

    Prototypicalnetworksforfew-shotlearning

    JakeSnell, KevinSwersky, andRichardZemel. Prototypicalnetworksforfew-shotlearning. Advances in neural information processing systems, 30, 2017. 21

  23. [31]

    Test- time training with self-supervision for generalization under distribution shifts

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test- time training with self-supervision for generalization under distribution shifts. InInterna- tional conference on machine learning, pages 9229–9248. PMLR, 2020

  24. [32]

    Learning to compare: Relation network for few-shot learning

    Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip HS Torr, and Timothy M Hospedales. Learning to compare: Relation network for few-shot learning. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 1199–1208, 2018

  25. [33]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan. Efficientnet: Rethinking model scaling for convolutional neural networks. arXiv preprint arXiv:1905.11946, 2019

  26. [34]

    Foct: Few-shot industrial anomaly detection with foreground- aware online conditional transport

    Long Tian, Hongyi Zhao, Ruiying Lu, Rongrong Wang, YuJie Wu, Liming Wang, Xiong- peng He, and Xiyang Liu. Foct: Few-shot industrial anomaly detection with foreground- aware online conditional transport. InACM Multimedia 2024

  27. [35]

    Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection

    Chengjie Wang, Wenbing Zhu, Bin-Bin Gao, Zhenye Gan, Jiangning Zhang, Zhihao Gu, Shuguang Qian, Mingang Chen, and Lizhuang Ma. Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. InProceedings of the IEEE/CVF Conference on Compute...

  28. [36]

    Generalizing from a few examples: A survey on few-shot learning.ACM computing surveys (csur), 53(3):1–34, 2020

    Yaqing Wang, Quanming Yao, James T Kwok, and Lionel M Ni. Generalizing from a few examples: A survey on few-shot learning.ACM computing surveys (csur), 53(3):1–34, 2020

  29. [37]

    Learning unsuper- vised metaformer for anomaly detection

    Jhih-Ciang Wu, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Learning unsuper- vised metaformer for anomaly detection. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4369–4378, 2021

  30. [38]

    Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise

    Julian Wyatt, Adam Leach, Sebastian M Schmon, and Chris G Willcocks. Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 650–656, 2022

  31. [39]

    Pushing the limits of fewshot anomaly detection in industry vision: Graphcore.ICLR, 2023

    Guoyang Xie, Jingbao Wang, Jiaqi Liu, Feng Zheng, and Yaochu Jin. Pushing the limits of fewshot anomaly detection in industry vision: Graphcore.ICLR, 2023

  32. [40]

    A uni- fied model for multi-class anomaly detection.Advances in Neural Information Processing Systems, 35:4571–4584, 2022

    Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A uni- fied model for multi-class anomaly detection.Advances in Neural Information Processing Systems, 35:4571–4584, 2022

  33. [41]

    Wide residual networks.arXiv preprint arXiv:1605.07146, 2016

    Sergey Zagoruyko and Nikos Komodakis. Wide residual networks.arXiv preprint arXiv:1605.07146, 2016

  34. [42]

    Omnial: A unified cnn framework for unsupervised anomaly localization

    Ying Zhao. Omnial: A unified cnn framework for unsupervised anomaly localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3924–3933, 2023

  35. [43]

    Spot- the-difference self-supervised pre-training for anomaly detection and segmentation

    Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot- the-difference self-supervised pre-training for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392–408. Springer, 2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.