Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Multimodal Task Representation Memory Bank vs. Catastrophic Forgetting in Anomaly Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that a per-task multimodal memory of keys, prompts, and fused text-image features, sharpened by structure-based contrastive learning, outperforms existing unsupervised continual anomaly detectors and reduces catastrophic…

desk verdict Genuinely novel combination for unsupervised continual AD with meaningful forgetting-rate gains, but Eq. (2) as written inverts key retrieval and the paper lacks reproducibility. read the letter →

arxiv 2502.06194 v1 pith:EXGIKTSK submitted 2025-02-10 cs.CV

You Zhou , Jiangshan Zhao , Deyu Zeng , Zuo Zuo , Weixiang Liu , Zongze Wu This is my paper · ORCID
classification cs.CV
keywords unsupervisedcontinualanomalydetectioncatastrophicforgettingmultimodalpromptlearningmemorybankcontrastivestructure-awarerepresentationMVTecADVisA
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

Unsupervised continual anomaly detection asks a model to keep learning the normal appearance of new product classes without forgetting earlier ones, and never seeing a defect. The paper claims that a compact per-task memory—a key, a learnable prompt, and fused image-text knowledge—solves the forgetting problem better than existing continual anomaly detectors. The proposed Multimodal Task Representation Memory Bank (MTRMB) builds this memory with a Key-Prompt-Multimodal Knowledge (KPMK) mechanism that cross-attends BERT text features with vision transformer (ViT) image features, and sharpens it with Refined Structure-Based Contrastive Learning (RSCL), which uses an open-set detector and segmenter (Grounding DINO and SAM) to pull patches of the same structure together and push different structures apart. On MVTec AD and VisA, MTRMB reports average image-level AUROC of 0.941 and 0.900, an average detection accuracy of 0.921 at the lowest forgetting rate among compared methods, and average forgetting measures of 0.016 and 0.010. If these results hold, multimodal prompting plus structure-aware contrastive learning is a viable route to continual anomaly detection without replaying old data.

What carries the argument

The central object is the Multimodal Task Representation Memory Bank (MTRMB): a growing set of per-task tuples {key, prompt, multimodal knowledge} that together hold the model's memory of all previously seen normal classes. The key is a condensed middle-layer ViT feature vector used to retrieve the right task by cosine similarity; the prompt is a short learnable prefix inserted into the frozen ViT to adapt it to that task; the multimodal knowledge is the fused BERT/ViT feature set used for anomaly scoring. The mechanism that carries the argument is the pairing of KPMK's cross-modal attention—text-as-query/image-as-key and image-as-query/text-as-key—with RSCL's masked contrastive loss, which pulls patch features sharing a pseudo-structure label together and pushes different structures apart. Inference then uses a query function to select the best-matching key and computes patch-level Euclidean distances to the stored normal features to build the anomaly map.

What would settle it

Compute the agreement (for example, IoU) between the Grounding DINO plus SAM masks and manually annotated object or part structure on a sample of MVTec AD and VisA images, then re-run the RSCL ablation with the masks replaced by random or superpixel partitions. If agreement is near chance, or if random masks reproduce the 0.030 Image-AUC gain attributed to RSCL, the structure-based explanation is not doing the claimed work.

Watch

Extended reading notes

Core claim

The paper's central claim is that the catastrophic forgetting seen in memory-based anomaly detectors comes from incomplete single-modality task representations, and that storing each task as a {key, prompt, multimodal knowledge} tuple corrects this. During training, each task contributes a key condensed from middle-layer ViT features, a learnable prefix prompt inserted into a frozen vision transformer, and a multimodal knowledge set obtained by cross-attending text and image features. At inference, the test image queries the task key, retrieves the matching prompt, and scores anomalies by nearest-neighbour patch distance against the stored normal features, following the PatchCore-style memory matching procedure. RSCL makes the stored memory more compact by using Grounding DINO with SAM to create pseudo-segmentation masks, then applying a masked contrastive loss that groups patches within the same structural region and separates patches from different regions. The paper reports that this design achieves average detection accuracy 0.921 at the lowest forgetting rate, with average Image AUROC 0.941 on MVTec AD and 0.900 on VisA.

Load-bearing premise

The load-bearing premise is that Grounding DINO and SAM produce accurate structural masks for every normal training image; if the masks are noisy, miss the object, or mostly mark similar backgrounds, the contrastive loss reshapes features around the wrong structure and the reported gains are not real improvements.

Editorial extensions

If this is right

  • On MVTec AD, MTRMB reports Image AUROC 0.941 and average forgetting 0.016, compared with UCAD's 0.926 and 0.056, so old-task knowledge survives better as new classes arrive.
  • On VisA, MTRMB reports Image AUROC 0.900 and average forgetting 0.010, compared with UCAD's 0.874 and 0.039 and DNE's 0.610 and 0.179.
  • Ablations attribute +0.030 Image AUC and +0.021 Pixel AP to RSCL; KPMK adds +0.011 on both metrics, and Grounding DINO adds +0.006 Image AUC and +0.01 Pixel AP, so each component contributes to the reported gain.
  • Prompt length matters: length 5 gives the best trade-off in the paper's ablation, with Image AUC 0.941 versus 0.936, 0.931, and 0.925 for lengths 1, 10, and 15.
  • Because only keys, prompts, and normal knowledge are stored rather than raw images, the method does not need replay of old data, which fits privacy-sensitive industrial inspection settings.

Reading between the lines

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

  • An extension the paper leaves implicit: replacing Grounding DINO and SAM with cheaper region proposals (superpixels or random partitions) and checking whether the RSCL gain survives would show whether semantic structure or mere contrastive grouping drives the improvement.
  • A quantitative mask-quality audit—IoU between the pseudo-masks and human object or part annotations—would turn the paper's visual mask examples into a testable claim, since the authors only show qualitative examples.
  • The same architecture could be ported to other dense prediction tasks with per-task domains, such as video anomaly detection or multi-site medical imaging, where a task key plus prompt could select the right normal-knowledge memory at test time.
  • The memory bank grows with the number of tasks, and the paper does not analyze storage growth or query latency; a practical follow-up would measure memory use and retrieval time as classes accumulate.
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 / 5 minor

Summary. The paper proposes the Multimodal Task Representation Memory Bank (MTRMB) for unsupervised continual anomaly detection (UCAD). The method combines a Key-Prompt-Multimodal Knowledge (KPMK) mechanism, which stores task keys, learnable prompts, and multimodal feature knowledge, with Refined Structure-based Contrastive Learning (RSCL), which uses Grounding DINO and SAM to produce structure masks that drive contrastive learning. Experiments on MVTec AD and VisA report average Image AUROC of 0.941 and 0.900, with forgetting measures of 0.016 and 0.010, outperforming UCAD and other baselines, and the paper also introduces a continual anomaly detection benchmark of 15 methods.

Significance. If the results hold, the paper would make a useful contribution to UCAD: it is, to my knowledge, the first proposal to combine multimodal prompt-based continual learning with a task representation memory bank for unsupervised anomaly detection, and the benchmark of 15 methods is a valuable resource. The paper also includes ablation studies that isolate the contributions of KPMK, RSCL, and Grounding DINO. However, the central claim is currently supported mainly by single-run average numbers with no error bars, no code release, and an internal inconsistency in the key-query equation, so the significance is conditional on verification. No machine-checked proofs or reproducible code are shipped with the manuscript.

major comments (4)
  1. [§3.1, Eq. (2)] Equation (2) defines the best-matching key selection as Lkp = arg min Σ γ(q(x), Ki), where γ is the cosine similarity. Since cosine similarity is larger for more similar features, argmin selects the least similar key, which contradicts the stated goal of retrieving the best-matching prompt. This is not a stylistic issue: the same query mechanism is used at inference in Section 3.3, and Lkp enters the total loss in Eq. (8). If the implementation actually uses argmin, the reported Image AUROC values in Tables 1 and 3 are not reproducible from the described algorithm; if it uses argmax, the paper mis-specifies its core mechanism. The authors must correct Eq. (2) and provide evidence, such as task-retrieval accuracy, that the key query retrieves the intended task.
  2. [§4.2, Tables 1 and 3] The abstract and Section 4.2 claim that MTRMB 'significantly outperforms' state-of-the-art methods, but the margins over UCAD are 0.015 Image AUROC on MVTec AD and 0.026 on VisA. No error bars, multiple-seed results, or significance tests are reported, so these small margins may be within run-to-run variability. The paper also does not specify the task order, optimizer, learning rate, number of epochs, or random seeds used in the continual learning protocol, which makes it impossible to assess whether the comparison is fair and reproducible. I request that the authors report mean and standard deviation over at least three seeds and provide the exact training protocol.
  3. [§3.2, Eqs. (5)-(6) and Figure 4] RSCL is a central contribution and the ablation in Table 4 attributes a 0.03 Image AUC gain to it. This gain depends on the quality of the pseudo-labels Y generated by Grounding DINO and SAM. The paper provides only visual examples in Figure 4 and no quantitative mask-quality metric, nor a comparison against raw SAM masks or an analysis of what happens when the masks are noisy. The paper's own t-SNE discussion in Figure 3 notes that background regions are similar across classes, so the contrastive pull/push for background patches may be based on uninformative structure. I recommend adding a quantitative evaluation of mask agreement (e.g., IoU with human annotations or with the anomaly-free ground-truth regions) and a sensitivity analysis over mask-quality thresholds.
  4. [§4.1 and §4.3] The experimental setup is incomplete for a claim of state-of-the-art performance. Section 4.1 does not report the pretrained backbone details beyond 'the fifth layer of the pretrained vision transformer', the text encoder setup for BERT, the prompt initialization, the temperature τ, the loss weight λ, or the number of training epochs. Table 5 ablates prompt length, but not τ or λ. Since these hyperparameters directly affect the KPMK and RSCL losses, the sensitivity of the reported averages to these choices should be documented.
minor comments (5)
  1. [Throughout] There are multiple typos and inconsistent names: 'MVtec' should be 'MVTec', 'Visa' should be 'VisA', 'corrsponding' should be 'corresponding', 'P atchcore' should be 'PatchCore', and the conclusion uses 'RSML' where the method is called 'RSCL' elsewhere.
  2. [§3.2] The sentence 'As illustrated in Figure ??' contains an unresolved figure reference; the manuscript should point to the actual figure showing Grounding DINO and SAM masks.
  3. [Eq. (12)] The definition of avgFM is difficult to parse as typeset: the expression 'max l∈{1,...,k−1} T l,j − T k,j' should clarify whether the maximum is taken over previous tasks l and whether T_{l,j} and T_{k,j} are the accuracies on task j after training tasks l and k, respectively.
  4. [§4.2, Figure 5] The qualitative comparison in Figure 5 references AdaClip and WinClip, but these methods are not included in the quantitative benchmark tables and are not described in the experimental setup; please add appropriate details or citations.
  5. [§4.3, Table 4] The ablation table reports a single average over MVTec AD and VisA; it would be clearer to report per-dataset results as well, since the two datasets have different class counts and difficulty levels.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: KPMK and RSCL are built from external features and normal-training data, and the claimed results are evaluated on held-out benchmarks.

full rationale

The derivation chain is self-contained. KPMK constructs task keys from ViT features of normal training images and selects prompts by cosine similarity; RSCL obtains structure pseudo-labels from external Grounding DINO and SAM and computes contrastive losses on those masks (Eqs. 5-7). The memory bank is populated from normal training data and evaluated against held-out test images of MVTec AD and VisA, and all comparisons are against external baselines rather than quantities fitted from the evaluation targets. The only author-overlapping citation is [55], a general survey of anomaly detection; it is used as background and is not load-bearing for the method's central claims. The internal inconsistency in Eq. 2 (Lkp written as an argmin over cosine similarity, which is dimensionally and semantically mismatched with its use as a loss in Eq. 8) is a correctness and rigor concern, not a circularity, because the reported metric does not reduce to the loss by construction. Likewise, hyperparameters selected on the benchmark test sets are an overfitting concern, not a circularity. No self-definitional, fitted-input-as-prediction, self-citation-load-bearing, or uniqueness-imported-from-authors step is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The central claim rests on a handful of unstated or weakly justified choices: the pseudo-label quality of Grounding DINO/SAM masks, the availability of class-name text for each task, the choice of ViT layer 5 for keys and anomaly features, and unspecified hyperparameters (prompt length, temperature, loss weight). These are not derived from theory and are not independently validated outside the reported tables.

free parameters (5)
  • Prompt length L_p = 5
    Ablation in Table 5 selects prompt length 5 as best on the combined MVTec/VisA metrics; this is tuned on the evaluation datasets rather than fixed by theory.
  • Temperature coefficient tau = not specified
    Appears in contrastive losses (Eqs. 5 and 7) but no value or tuning procedure is reported; central training depends on it.
  • Loss weight lambda for key loss = not specified
    Total loss L_all = L_contra + L_cross + lambda*L_kp, but lambda is not specified or ablated.
  • Feature extraction layer o = 5
    Task keys and anomaly features are taken from layer 5 of ViT (Sections 3.1 and 3.3); chosen by heuristic that middle layers contain contextual information, with no supporting layer comparison.
  • Number of keys/prompts N and FPS condensation budget = not specified
    Memory bank capacity and farthest-point-sampling key count are not reported, although they determine capacity and compute.
assumptions (4)
  • domain assumption Grounding DINO and SAM pseudo-labels Y accurately segment meaningful structural regions in every normal training image of MVTec AD and VisA classes.
    RSCL's contrastive loss (Eq. 6) pulls and pushes patches based on Y; if masks are noisy or miss the object, the learned structure is wrong. The paper gives only visual examples (Figure 4), no mask accuracy metric.
  • domain assumption Text class names such as 'Metal nut' are available for each task and provide complementary signal to images.
    KPMK's cross-modal attention (Eqs. 3-4) and L_cross (Eq. 7) require text features; in a scenario without class names, the multimodal component cannot be used.
  • ad hoc to paper The fifth-layer ViT features contain sufficient task-discriminative information for key matching and anomaly scoring.
    Layer o=5 is chosen by heuristic in Section 3.1 and used for both keys and anomaly features; no comparison across layers is provided.
  • domain assumption Pre-trained frozen ViT/BERT embeddings generalize across the continuous task sequence without adaptation.
    Backbones are frozen; only prompts and small heads are tuned. If feature shift across tasks is large, memory bank similarity will degrade. This is standard in prompt-based continual learning but not validated here.
invented entities (2)
  • Key-Prompt-Multimodal Knowledge tuple {K, P, MK}
    purpose: Per-task memory storing a task key, learnable prompt, and multimodal normal knowledge to enable task retrieval and knowledge transfer in UCAD.
    This is a new architectural component introduced by the paper; it has no falsifiable signature outside the reported benchmark, and its effectiveness is only evidenced by the paper's own tables.
  • Task key condensation via FPS
    purpose: Compress layer-5 features into a small key set for matching test images to the correct task prompt.
    The FPS-based key set is a design invention with no external validation; no analysis of key retrieval accuracy is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Task Representation Memory Bank vs. Catastrophic Forgetting in Anomaly Detection." pith.science (2026). https://pith.science/paper/EXGIKTSK

@misc{pith2026250206194,
  author       = {Pith},
  title        = {Pith review of: Multimodal Task Representation Memory Bank vs. Catastrophic Forgetting in Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EXGIKTSK}},
  note         = {Machine review of arXiv:2502.06194}
}
read the original abstract

Unsupervised Continuous Anomaly Detection (UCAD) faces significant challenges in multi-task representation learning, with existing methods suffering from incomplete representation and catastrophic forgetting. Unlike supervised models, unsupervised scenarios lack prior information, making it difficult to effectively distinguish redundant and complementary multimodal features. To address this, we propose the Multimodal Task Representation Memory Bank (MTRMB) method through two key technical innovations: A Key-Prompt-Multimodal Knowledge (KPMK) mechanism that uses concise key prompts to guide cross-modal feature interaction between BERT and ViT. Refined Structure-based Contrastive Learning (RSCL) leveraging Grounding DINO and SAM to generate precise segmentation masks, pulling features of the same structural region closer while pushing different structural regions apart. Experiments on MVtec AD and VisA datasets demonstrate MTRMB's superiority, achieving an average detection accuracy of 0.921 at the lowest forgetting rate, significantly outperforming state-of-the-art methods. We plan to open source on GitHub.

Figures

Figures reproduced from arXiv: 2502.06194 by the authors.

Figure 1
Figure 1. Comparison between UCAD with a memory bank and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of UCAD using multimodal Task Representation Memory Bank. (a) Text-image data is input during the training [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. The first row shows the test images, the second row shows [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Visualization examples of continual anomaly detection. The first row displays the original anomaly images, the second row shows [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. DeCoFlow: Structural Decomposition of Normalizing Flows for Continual Anomaly Detection

    cs.CV 2026-06 unverdicted novelty 6.5 of 10

    DeCoFlow decomposes normalizing flow subnets into frozen bases and low-rank adapters with alignment, auxiliary layers, and tail-aware loss to achieve continual anomaly detection with zero forgetting and few added parameters.

Reference graph

Works this paper leans on

56 extracted references · 41 canonical work pages · cited by 1 Pith paper

  1. [10]

    Towards continual adapta- tion in industrial anomaly detection,

    W. Li, J. Zhan, J. Wang, B. Xia, B.-B. Gao, J. Liu, C. Wang, and F. Zheng, “Towards continual adapta- tion in industrial anomaly detection,” in Proceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 2871–2880

    work page 2022

  2. [1]

    Mvtec ad–a comprehensive real-world dataset for un- supervised anomaly detection,

    P. Bergmann, M. Fauser, D. Sattlegger, and C. Steger, “Mvtec ad–a comprehensive real-world dataset for un- supervised anomaly detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 9592–9600

    work page 2019

  3. [2]

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

    Y . Zou, J. Jeong, L. Pemula, D. Zhang, and O. Dabeer, “Spot-the-difference self-supervised pre-training for anomaly detection and segmentation,” arXiv preprint arXiv:2207.14315, 2022

    arXiv 2022

  4. [3]

    Deep industrial image anomaly detection: A survey,

    J. Liu, G. Xie, J. Wang, S. Li, C. Wang, F. Zheng, and Y . Jin, “Deep industrial image anomaly detection: A survey,” Machine Intelligence Research, vol. 21, no. 1, pp. 104–135, 2024

    2024

  5. [4]

    Im-iad: Industrial image anomaly detection benchmark in manufacturing,

    G. Xie, J. Wang, J. Liu, J. Lyu, Y . Liu, C. Wang, F. Zheng, and Y . Jin, “Im-iad: Industrial image anomaly detection benchmark in manufacturing,”IEEE Transac- tions on Cybernetics, 2024

    work page 2024

  6. [5]

    Cfa: Coupled- hypersphere-based feature adaptation for target- oriented anomaly localization,

    S. Lee, S. Lee, and B. C. Song, “Cfa: Coupled- hypersphere-based feature adaptation for target- oriented anomaly localization,” IEEE Access , vol. 10, pp. 78 446–78 454, 2022

    work page 2022

  7. [6]

    Cflow-ad: Real-time unsupervised anomaly detection with local- ization via conditional normalizing flows,

    D. Gudovskiy, S. Ishizaka, and K. Kozuka, “Cflow-ad: Real-time unsupervised anomaly detection with local- ization via conditional normalizing flows,” in Proceed- ings of the IEEE/CVF winter conference on applications of computer vision, 2022, pp. 98–107

    work page 2022

  8. [7]

    Fully convolutional cross-scale-flows for image-based defect detection,

    M. Rudolph, T. Wehrbein, B. Rosenhahn, and B. Wandt, “Fully convolutional cross-scale-flows for image-based defect detection,” inProceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision , 2022, pp. 1088–1097

    work page 2022

Show all 56 references
  1. [8]

    Cutpaste: Self-supervised learning for anomaly detection and lo- calization,

    C.-L. Li, K. Sohn, J. Yoon, and T. Pfister, “Cutpaste: Self-supervised learning for anomaly detection and lo- calization,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 9664–9674

    2021

  2. [9]

    Iddm: An incremental dual- network detection model for in-situ inspection of large- scale complex product,

    F. Zhang and Z. Chen, “Iddm: An incremental dual- network detection model for in-situ inspection of large- scale complex product,” Journal of Industrial Informa- tion Integration, vol. 33, p. 100463, 2023

    2023

  3. [11]

    Cross-modal alternating learn- ing with task-aware representations for continual learn- ing,

    W. Li, B.-B. Gao, B. Xia, J. Wang, J. Liu, Y . Liu, C. Wang, and F. Zheng, “Cross-modal alternating learn- ing with task-aware representations for continual learn- ing,” IEEE Transactions on Multimedia, 2023

    2023

  4. [12]

    Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection,

    X. Zhang, M. Xu, and X. Zhou, “Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2024, pp. 16 699–16 708

    2024

  5. [13]

    Towards total recall in industrial anomaly detection,

    K. Roth, L. Pemula, J. Zepeda, B. Sch ¨olkopf, T. Brox, and P. Gehler, “Towards total recall in industrial anomaly detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 14 318–14 328

    2022

  6. [14]

    Uninformed students: Student-teacher anomaly detec- tion with discriminative latent embeddings,

    P. Bergmann, M. Fauser, D. Sattlegger, and C. Steger, “Uninformed students: Student-teacher anomaly detec- tion with discriminative latent embeddings,” in Pro- ceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, 2020, pp. 4183–4192

    2020

  7. [15]

    Multiresolution knowledge distil- lation for anomaly detection,

    M. Salehi, N. Sadjadi, S. Baselizadeh, M. H. Rohban, and H. R. Rabiee, “Multiresolution knowledge distil- lation for anomaly detection,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 14 902–14 912

    2021

  8. [16]

    Anomaly detection via reverse dis- tillation from one-class embedding,

    H. Deng and X. Li, “Anomaly detection via reverse dis- tillation from one-class embedding,” in Proceedings of the IEEE/CVF conference on computer vision and pat- tern recognition, 2022, pp. 9737–9746

    2022

  9. [17]

    Revis- iting reverse distillation for anomaly detection,

    T. D. Tien, A. T. Nguyen, N. H. Tran, T. D. Huy, S. Duong, C. D. T. Nguyen, and S. Q. Truong, “Revis- iting reverse distillation for anomaly detection,” in Pro- ceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, 2023, pp. 24 511–24 520

    2023

  10. [18]

    Efficientad: Ac- curate visual anomaly detection at millisecond-level la- tencies,

    K. Batzner, L. Heckler, and R. K¨onig, “Efficientad: Ac- curate visual anomaly detection at millisecond-level la- tencies,” in Proceedings of the IEEE/CVF Winter Con- ference on Applications of Computer Vision , 2024, pp. 128–138

    2024

  11. [19]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su et al., “Grounding dino: Marrying dino with grounded pre-training for open-set object detection,” inEuropean Conference on Computer Vision. Springer, 2025, pp. 38–55

    2025

  12. [20]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.- Y . Loet al., “Segment anything,” in Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, 2023, pp. 4015–4026

    2023

  13. [21]

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

    J. Jeong, Y . Zou, T. Kim, D. Zhang, A. Ravichan- dran, and O. Dabeer, “Winclip: Zero-/few-shot anomaly classification and segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 19 606–19 616

    2023

  14. [22]

    Adaclip: Adapting clip with hybrid learn- able prompts for zero-shot anomaly detection,

    Y . Cao, J. Zhang, L. Frittoli, Y . Cheng, W. Shen, and G. Boracchi, “Adaclip: Adapting clip with hybrid learn- able prompts for zero-shot anomaly detection,” inEuro- pean Conference on Computer Vision. Springer, 2025, pp. 55–72

    2025

  15. [23]

    X. Chen, Y . Han, and J. Zhang, “April-gan: A zero-/few- shot anomaly classification and segmentation method for cvpr 2023 vand workshop challenge tracks 1&2: 1st place on zero-shot ad and 4th place on few-shot ad,” arXiv preprint arXiv:2305.17382, 2023

    2023 arXiv

  16. [24]

    Winning solution for the cvpr2023 visual anomaly and novelty detection challenge: Multimodal prompt- ing for data-centric anomaly detection,

    Y . Cao, X. Xu, C. Sun, Y . Cheng, L. Gao, and W. Shen, “Winning solution for the cvpr2023 visual anomaly and novelty detection challenge: Multimodal prompt- ing for data-centric anomaly detection,” arXiv preprint arXiv:2306.09067, 2023

    2023 arXiv

  17. [25]

    Dinov2: Learning ro- bust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby et al. , “Dinov2: Learning ro- bust visual features without supervision,”arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  18. [26]

    Patch svdd: Patch-level svdd for anomaly detection and segmentation,

    J. Yi and S. Yoon, “Patch svdd: Patch-level svdd for anomaly detection and segmentation,” in Proceedings of the Asian conference on computer vision, 2020

    2020

  19. [27]

    Promptad: Learning prompts with only normal samples for few-shot anomaly detection,

    X. Li, Z. Zhang, X. Tan, C. Chen, Y . Qu, Y . Xie, and L. Ma, “Promptad: Learning prompts with only normal samples for few-shot anomaly detection,” in Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 838–16 848

    2024

  20. [28]

    Mocca: Multilayer one-class classification for anomaly detection,

    F. V . Massoli, F. Falchi, A. Kantarci, S ¸. Akti, H. K. Ekenel, and G. Amato, “Mocca: Multilayer one-class classification for anomaly detection,”IEEE transactions on neural networks and learning systems, vol. 33, no. 6, pp. 2313–2323, 2021

    2021

  21. [29]

    Simplenet: A simple network for image anomaly detection and local- ization,

    Z. Liu, Y . Zhou, Y . Xu, and Z. Wang, “Simplenet: A simple network for image anomaly detection and local- ization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 20 402–20 411

    2023

  22. [30]

    Anomaly detection under distribution shift,

    T. Cao, J. Zhu, and G. Pang, “Anomaly detection under distribution shift,” in Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision , 2023, pp. 6511–6523

    2023

  23. [31]

    Ms- flow: Multiscale flow-based framework for unsuper- vised anomaly detection,

    Y . Zhou, X. Xu, J. Song, F. Shen, and H. T. Shen, “Ms- flow: Multiscale flow-based framework for unsuper- vised anomaly detection,”IEEE Transactions on Neural Networks and Learning Systems, 2024

    2024

  24. [32]

    Push- ing the limits of fewshot anomaly detection in industry vision: Graphcore,

    G. Xie, J. Wang, J. Liu, F. Zheng, and Y . Jin, “Push- ing the limits of fewshot anomaly detection in industry vision: Graphcore,” arXiv preprint arXiv:2301.12082 , 2023

    2023 arXiv

  25. [33]

    A unified model for multi-class anomaly detection,

    Z. You, L. Cui, Y . Shen, K. Yang, X. Lu, Y . Zheng, and X. Le, “A unified model for multi-class anomaly detection,” Advances in Neural Information Processing Systems, vol. 35, pp. 4571–4584, 2022

    2022

  26. [34]

    Omnial: A unified cnn framework for unsu- pervised anomaly localization,

    Y . Zhao, “Omnial: A unified cnn framework for unsu- pervised anomaly localization,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3924–3933

    2023

  27. [35]

    Long- tailed anomaly detection with learnable class names,

    C.-H. Ho, K.-C. Peng, and N. Vasconcelos, “Long- tailed anomaly detection with learnable class names,” in Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, 2024, pp. 12 435– 12 446

    2024

  28. [36]

    Natural synthetic anomalies for self-supervised anomaly detec- tion and localization,

    H. M. Schl ¨uter, J. Tan, B. Hou, and B. Kainz, “Natural synthetic anomalies for self-supervised anomaly detec- tion and localization,” inEuropean Conference on Com- puter Vision. Springer, 2022, pp. 474–489

    2022

  29. [37]

    Multimodal parameter-efficient few-shot class incremental learning,

    M. D’Alessandro, A. Alonso, E. Calabr ´es, and M. Galar, “Multimodal parameter-efficient few-shot class incremental learning,” in Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, 2023, pp. 3393–3403

    2023

  30. [38]

    Draem- a discriminatively trained reconstruction embedding for surface anomaly detection,

    V . Zavrtanik, M. Kristan, and D. Sko ˇcaj, “Draem- a discriminatively trained reconstruction embedding for surface anomaly detection,” in Proceedings of the IEEE/CVF international conference on computer vi- sion, 2021, pp. 8330–8339

    2021

  31. [39]

    In- dustrial surface defect detection and localization us- ing multi-scale information focusing and enhancement ganomaly,

    J. Peng, H. Shao, Y . Xiao, B. Cai, and B. Liu, “In- dustrial surface defect detection and localization us- ing multi-scale information focusing and enhancement ganomaly,” Expert Systems with Applications, vol. 238, p. 122361, 2024

    2024

  32. [40]

    Prior normality prompt transformer for multi-class industrial image anomaly detection,

    H. Yao, Y . Cao, W. Luo, W. Zhang, W. Yu, and W. Shen, “Prior normality prompt transformer for multi-class industrial image anomaly detection,” arXiv preprint arXiv:2406.11507, 2024

    2024 arXiv

  33. [41]

    Unsupervised continual anomaly detection with contrastively-learned prompt,

    J. Liu, K. Wu, Q. Nie, Y . Chen, B.-B. Gao, Y . Liu, J. Wang, C. Wang, and F. Zheng, “Unsupervised continual anomaly detection with contrastively-learned prompt,” in Proceedings of the AAAI Conference on Ar- tificial Intelligence, vol. 38, no. 4, 2024, pp. 3639–3647

    2024

  34. [42]

    Riemannian walk for incremental learning: Un- derstanding forgetting and intransigence,

    A. Chaudhry, P. K. Dokania, T. Ajanthan, and P. H. Torr, “Riemannian walk for incremental learning: Un- derstanding forgetting and intransigence,” in Proceed- ings of the European conference on computer vision (ECCV), 2018, pp. 532–547

    2018

  35. [43]

    Imagenet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition. Ieee, 2009, pp. 248–255

    2009

  36. [44]

    Anomaly localization by modeling perceptual features,

    D. Dehaene and P. Eline, “Anomaly localization by modeling perceptual features,” arXiv preprint arXiv:2008.05369, 2020

    2008 arXiv

  37. [45]

    Sub-image anomaly de- tection with deep pyramid correspondences,

    N. Cohen and Y . Hoshen, “Sub-image anomaly de- tection with deep pyramid correspondences,” arXiv preprint arXiv:2005.02357, 2020

    2005 arXiv

  38. [46]

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

    T. Defard, A. Setkov, A. Loesch, and R. Audigier, “Padim: a patch distribution modeling framework for anomaly detection and localization,” in International Conference on Pattern Recognition . Springer, 2021, pp. 475–489

    2021

  39. [47]

    Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows,

    J. Yu, Y . Zheng, X. Wang, W. Li, Y . Wu, R. Zhao, and L. Wu, “Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows,” arXiv preprint arXiv:2111.07677, 2021

    2021 arXiv

  40. [48]

    Reconpatch: Contrastive patch representation learning for industrial anomaly detection,

    J. Hyun, S. Kim, G. Jeon, S. H. Kim, K. Bae, and B. J. Kang, “Reconpatch: Contrastive patch representation learning for industrial anomaly detection,” in Proceed- ings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, 2024, pp. 2052–2061

    2024

  41. [49]

    Adam: A method for stochastic opti- mization,

    D. P. Kingma, “Adam: A method for stochastic opti- mization,” arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  42. [50]

    Dual- prompt: Complementary prompting for rehearsal-free continual learning,

    Z. Wang, Z. Zhang, S. Ebrahimi, R. Sun, H. Zhang, C.- Y . Lee, X. Ren, G. Su, V . Perot, J. Dy et al. , “Dual- prompt: Complementary prompting for rehearsal-free continual learning,” in European Conference on Com- puter Vision. Springer, 2022, pp. 631–648

    2022

  43. [51]

    Learning to prompt for continual learning,

    Z. Wang, Z. Zhang, C.-Y . Lee, H. Zhang, R. Sun, X. Ren, G. Su, V . Perot, J. Dy, and T. Pfister, “Learning to prompt for continual learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 139–149

    2022

  44. [52]

    Visual prompt tuning,

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim, “Visual prompt tuning,” in European Conference on Computer Vision. Springer, 2022, pp. 709–727

    2022

  45. [53]

    The farthest point strategy for progressive image sam- pling,

    Y . Eldar, M. Lindenbaum, M. Porat, and Y . Y . Zeevi, “The farthest point strategy for progressive image sam- pling,” IEEE transactions on image processing , vol. 6, no. 9, pp. 1305–1315, 1997

    1997

  46. [54]

    Class-incremental learning: A sur- vey,

    D.-W. Zhou, Q.-W. Wang, Z.-H. Qi, H.-J. Ye, D.-C. Zhan, and Z. Liu, “Class-incremental learning: A sur- vey,” IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 2024

    2024

  47. [55]

    A sur- vey on rgb, 3d, and multimodal approaches for unsu- pervised industrial anomaly detection,

    Y . Lin, Y . Chang, X. Tong, J. Yu, A. Liotta, G. Huang, W. Song, D. Zeng, Z. Wu, Y . Wang et al. , “A sur- vey on rgb, 3d, and multimodal approaches for unsu- pervised industrial anomaly detection,” arXiv preprint arXiv:2410.21982, 2024

    arXiv

  48. [56]

    Recent advances of multimodal contin- ual learning: A comprehensive survey,

    D. Yu, X. Zhang, Y . Chen, A. Liu, Y . Zhang, P. S. Yu, and I. King, “Recent advances of multimodal contin- ual learning: A comprehensive survey,” arXiv preprint arXiv:2410.05352, 2024

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.