Pith. sign in

REVIEW 3 major objections 4 minor 149 references

Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that replacing class-specific modules with a learnable concept set and soft switching makes long-tailed anomaly detection work without class labels, beating class-aware methods and enabling online adaptation.

desk verdict A genuinely new benchmark and a plausible class-agnostic concept idea, but the 'no class labels' claim is undermined by a label-dependent sampler and the online detection headline is not in the tables. read the letter →

arxiv 2507.16946 v1 pith:55SAAPJM submitted 2025-07-22 cs.CV

classification cs.CV
keywords long-tailedanomalydetectiononlineclass-agnosticmodelsvision-languagevectorquantizationconceptlearningpromptcontinual
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

Long-tailed anomaly detection has so far relied on knowing which class each image belongs to, a label that is unavailable when a model must keep learning from an online stream. This paper tries to establish that class information is unnecessary: a relatively small set of learned concepts, found by majority vote over a vision-language vocabulary and refined during training, can stand in for the class set, with each image assigned a soft score over concepts. On that basis the authors build LTOAD, a class-agnostic reconstruction-plus-semantics model with concept-specific vector-quantized codebooks and normal/abnormal prompts, and pair it with an online algorithm (AAA) that up-weights samples whose pseudo anomaly maps look abnormal. They report that this pipeline beats the strongest class-aware offline baselines on several long-tailed benchmarks, including +4.63% image-AUROC on MVTec, and improves online segmentation in the most difficult stream configurations.

What carries the argument

The load-bearing object is the concept set Ĉ and the concept score p it induces. A vocabulary from a vision-language foundation model is scored against all training images, and the top bK vocabularies by majority vote initialize Ĉ; their text embeddings t_bc are then made learnable. At test time p = SoftMax({⟨f_f, t_bc⟩}) assigns each image a distribution over concepts, and every class-specific module is replaced by a p-weighted mixture over concept-specific modules. The reconstruction branch uses Concept-VQ, hierarchical vector-quantized autoencoders whose per-concept codebooks are initialized around t_bc and whose outputs are mixed by p; a prompt-conditioned generator synthesizes pseudo-normal and pseudo-abnormal features; and the semantics branch compares per-concept normal prompts with concept-specific abnormal prompts generated by a conversational AI. The online component, AAA, thresholds the model's own anomaly map into a pseudo abnormal mask and weights the per-sample loss by β when the reduction of bY exceeds τ, with EMA smoothing of updates.

What would settle it

Run LTOAD on a benchmark whose classes deliberately have near-zero semantic overlap with the foundation model's vocabulary, such as random out-of-domain nouns as concepts while holding the architecture fixed, and check whether its detection AUROC advantage over class-aware baselines disappears; the paper's Random out-domain experiment partially does this and shows degradation, so the strong version of the falsifier would sweep vocabulary size and overlap.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the class label c can be replaced by a soft label p over a learned concept set Ĉ, where bK = |Ĉ| does not need to equal K = |C|. Every class is assumed expressible as a composition of concepts, so the model's output is a p-weighted combination of concept-specific modules instead of a hard switch to the class-c module. The authors argue this is not merely a workaround for missing labels: because classes share concepts, the soft representation transfers knowledge across classes, handles unseen classes, and remains effective even when Ĉ is initialized from an out-of-domain vocabulary. Combined with the AAA update rule, the same class-agnostic model can be adapted on a stream of possibly abnormal images, which is the defining requirement of the new long-tailed online anomaly detection task they propose.

Load-bearing premise

The framework assumes that every anomaly-detection class can be represented as a weighted combination of concepts in the learned concept set, and that the majority-vote initialization over the foundation-model vocabulary actually surfaces those concepts from unlabeled training images alone.

Editorial extensions

If this is right

  • Offline long-tailed anomaly detection no longer requires class labels or the number of classes; the paper's experiments on MVTec, VisA, DAGM, and Uni-Medical show the class-agnostic model matching or exceeding class-aware baselines in most settings.
  • An online stream with unlabeled abnormal images can be exploited through the AAA update, which outperforms naive fine-tuning especially when head or tail classes arrive in early sessions.
  • Unseen classes that never appeared in training can still be scored, because the concept set is not tied to the training class set; cross-dataset experiments show adaptation without retraining for the new domain.
  • Because bK can be smaller than K, the model uses fewer concept-specific parameters and runs faster than the class-aware LTAD baseline it extends.

Reading between the lines

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

  • Extending beyond the paper, the same soft-switching recipe could be dropped into other class-aware anomaly detection architectures, not just reconstruction-semantics pipelines, if the composition assumption is the true source of the gain.
  • A sharper test of the concept mechanism would vary the semantic overlap between the foundation vocabulary and the target classes while holding everything else fixed; the paper's out-domain Random experiment already shows degradation, so one could measure whether performance scales with overlap.
  • AAA's design implies that online gains are concentrated in samples the model already suspects are abnormal; one could test whether β and τ need to be stream-adaptive to avoid overfitting when a session is dominated by one anomaly type.
  • The eight stream configurations of the proposed LTOAD benchmark define a reusable stress test for forgetting and overfitting in online anomaly detection, so future online methods can be compared directly on the released splits.
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

3 major / 4 minor

Summary. The paper introduces LTOAD, a long-tailed online anomaly detection task and benchmark, and proposes a class-agnostic framework based on a learnable concept set (\hat{C}) with soft switching, a Concept-VQ reconstruction branch, prompt-conditioned data augmentation, and an Anomaly Adaptive (AAA) online learning algorithm. It reports SOTA offline results on LTAD benchmarks (MVTec, VisA, DAGM, Uni-Medical) and online segmentation results on the new benchmark, and the abstract further claims an image-AUROC gain of +0.53% in the online setting.

Significance. If the claims hold, this is a meaningful contribution: a unified AD model that removes class-label requirements, supports unseen classes, and adapts online, with large and internally consistent offline gains across several datasets. The paper's strengths include the released benchmark, extensive per-class results, ablations of each module, and computational-cost comparisons. However, the verification of the online claim and the label-free claim requires addressing the issues below, so the significance is conditional on those fixes.

major comments (3)
  1. [Abstract / §5.2 / Tab. 4] The abstract's headline online result is "we achieve +0.53% image-AUROC compared to baselines," but the paper reports only pixel-level AUROC for anomaly segmentation in Tab. 4 and in Appendix Tables A51–A53. No online image-level AUROC table or curve is provided anywhere in the manuscript, so the +0.53% figure cannot be verified from the reported results; the average in Tab. 4 (95.28 for LTOAD/AAA) does not reproduce the claimed gain relative to any listed baseline.
  2. [Appendix A1.5] The statement "During training, we use the balance sampler for sampling long-tailed data distribution in LTAD [34]" contradicts the 'class-agnostic' definition in Tabs. 2–3 and the claim that LTOAD does not require class labels, because the LTAD balance sampler needs per-image class labels to rebalance batches. This makes the offline training procedure label-dependent, which weakens the comparison to class-aware methods and undermines the central class-agnostic claim. The paper should either replace the balance sampler with a label-free procedure or explicitly clarify how the sampler is used without class labels.
  3. [§2.2 / Appendix A1.6] The concept set is claimed to be discovered from unlabeled images by majority voting over the ALIGN vocabulary, but the listed concepts for each dataset (e.g., MVTec: semiconductor, zipper, beech, walnut, circuit, microscopy, mahogany, hardwood, medicines, antibiotics) are suspiciously aligned with the class semantics. The paper does not show the raw voting outputs or provide evidence that these lists are the automatic voting results rather than manually curated from class names. This matters because the 'no class labels' claim depends on the concept set being derived without class information; please provide the voting outputs or otherwise justify the procedure.
minor comments (4)
  1. [Alg. 1 / A1.5] Define T more precisely: T(\hat{Y}) = 0.95 · r(\hat{Y}) is a scalar, but §3.2 calls \hat{M} a pseudo abnormal map and refers to a mask-dependent loss; the relationship between the threshold function, the scalar threshold, and the loss mask is unclear.
  2. [Tab. 1 caption] The caption lists the ordering type as "head-first, head-first, else"; the second entry should presumably be "tail-first".
  3. [Tab. 7] The column order "Ctail Chead C" in Tab. 7 is inconsistent with the ordering used in other tables and in the surrounding text; please standardize it.
  4. [Tab. A1] Tab. A1 lists class elements but does not visually mark which classes belong to C_head vs. C_tail, although the preceding sentence says it does.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported anomaly-detection numbers are measured against external ground truth, and the concept set is a VLM-initialized latent structure, not a fitted target.

full rationale

The paper's central claims are empirical: LTOAD is evaluated on fixed benchmarks (MVTec, VisA, DAGM, Uni-Medical) with held-out ground-truth labels, so the headline AUROC gains are not constructed from the method's own inputs. The concept set bC is initialized by majority voting over ALIGN's vocabulary using only unlabeled training images (Sec. 2.2), and the Random Out-domain ablation in Sec. 5.3 shows that even an unrelated concept set remains competitive, so the method does not rely on bC secretly encoding the class labels. The AAA online algorithm does use the model's own predictions as pseudo-labels (Algorithm 1), but this is a standard self-training mechanism whose effect is evaluated on held-out data; it does not define the target metric by construction. The use of LTAD's balance sampler (Appendix A1.5) is a potential inconsistency with the 'class-agnostic' claim and is a correctness risk, but it is not circular: the reported numbers are not analytically equal to the sampler's output, and the comparison is against external baselines. Self-citations to LTAD [34] and HVQ [57] refer to published, externally reproducible work and are used as baselines or architectural starting points, not as load-bearing justifications for the paper's own predictions. No equation in the paper reduces a predicted quantity to a fitted parameter by construction, and no load-bearing uniqueness theorem or ansatz is imported from the authors' prior work. The suspicious overlap between the listed concepts (Appendix A1.6) and class semantics is an empirical observation about VLM alignment, and the paper itself tests the case where such overlap is absent (Random O-D), which counts against circularity. Overall, the derivation chain is self-contained and externally falsifiable.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

No new physical entities are postulated; the 'concepts' are learned latent representations. The central assumptions are about concept representability, vocabulary alignment, and the availability of abnormal images in the online stream. The listed hyperparameters are chosen by hand and are load-bearing for the AAA algorithm and the fusion of the two branches.

free parameters (6)
  • bK (number of concepts) = 10 for all datasets
    Chosen by hand; controls the size of the concept set and model capacity. No sensitivity analysis is provided.
  • alpha (fusion weight in Eq. 4) = 0.3
    Fixed weight between reconstruction and semantics branches; not swept.
  • beta (pseudo-abnormal sample weight in AAA) = 5
    Hyperparameter in the anomaly adaptive loss; no study of sensitivity.
  • tau (anomaly score threshold in AAA) = 0.2
    Hyperparameter in AAA; determines which images receive the higher loss weight.
  • gamma (EMA decay in AAA) = 0.3
    Hyperparameter in AAA for stable parameter updates.
  • threshold scale T(Y)=0.95*r(Y) = 0.95
    Constant in the pseudo-label threshold function; chosen without reported tuning.
assumptions (3)
  • domain assumption Every class can be represented as a composition of concepts in a learned concept set bC.
    Sec. 2.2 states this assumption; if concepts do not cover the semantics of the classes, the soft-label switching and concept-specific modules lose validity.
  • domain assumption The ALIGN text vocabulary contains words whose embeddings are semantically aligned with the visual appearance of the training classes.
    The majority-vote concept initialization in Sec. 2.2 depends on this; the Random out-domain experiment in Sec. 5.3 shows degradation when vocabulary is unrelated.
  • domain assumption The online stream contains a mix of normal and abnormal images, and the model can exploit unlabeled abnormal images via pseudo-labels.
    This is the operational assumption of the online benchmark in Sec. 3.1; it means the method effectively uses test-distribution data during adaptation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts." pith.science (2026). https://pith.science/paper/55SAAPJM

@misc{pith2026250716946,
  author       = {Pith},
  title        = {Pith review of: Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/55SAAPJM}},
  note         = {Machine review of arXiv:2507.16946}
}
read the original abstract

Anomaly detection (AD) identifies the defect regions of a given image. Recent works have studied AD, focusing on learning AD without abnormal images, with long-tailed distributed training data, and using a unified model for all classes. In addition, online AD learning has also been explored. In this work, we expand in both directions to a realistic setting by considering the novel task of long-tailed online AD (LTOAD). We first identified that the offline state-of-the-art LTAD methods cannot be directly applied to the online setting. Specifically, LTAD is class-aware, requiring class labels that are not available in the online setting. To address this challenge, we propose a class-agnostic framework for LTAD and then adapt it to our online learning setting. Our method outperforms the SOTA baselines in most offline LTAD settings, including both the industrial manufacturing and the medical domain. In particular, we observe +4.63% image-AUROC on MVTec even compared to methods that have access to class labels and the number of classes. In the most challenging long-tailed online setting, we achieve +0.53% image-AUROC compared to baselines. Our LTOAD benchmark is released here: https://doi.org/10.5281/zenodo.16283852 .

Figures

Figures reproduced from arXiv: 2507.16946 by the authors.

Figure 1
Figure 1. Comparison of LTOAD to class-aware anomaly de￾tection methods on offline and online learning. (a) Class-aware methods have a class-specific module for each class c in the class set C. (b) These methods cannot work in online learning when the class labels are unavailable. (c) LTOAD solves this prob￾lem by learning a concept set Cb to approximate C. We note that Kb = |C| b does not need to match K = |C|. (d) In online… view at source ↗
Figure 2
Figure 2. Proposed class-agnostic pipeline. We construct Cb, and the correspondent normal prompts P n and abnormal prompts P a . The concept score p is assigned to each image X by computing the similarity between X and each bc ∈ Cb. It then controls the soft switching mechanism in our class-agnostic reconstruction module R and semantics module S. R reconstructs F i into F r through Concept VQ and output Yb R by measuring the … view at source ↗
Figure 3
Figure 3. Qualitative comparison among LTAD [34], HVQ [57], and LTOAD on MVTec offline LTAD [34] exp100 benchmark. In￾puts from C head / C tail are outlined in blue / red. Smaller ground truth anomaly masks are circled in pink. Method CA Online B B-HF B-TF D2-HF D2-TF D5-HF D5-TF D5-M Avg. LTAD [34] ✗ ✗ 93.46 93.46 93.46 93.46 93.46 93.46 93.46 93.46 93.46 LTAD [34] ✗ AN 94.12 93.80 94.07 93.70 94.08 93.49 93.84 93.65 93.84 H… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Performance curve of A N and A AA in pixel-level AU￾ROC on C head and C tail. They are offline-trained on exp100 [34] and online-trained on D5-HF or D5-TF. same datasets as the offline benchmark [34], i.e., MVTec, VisA, and DAGM. We split the testing set into DO and DE…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

149 extracted references · 71 canonical work pages

  1. [34]

    Long-tailed anomaly detection with learnable class names

    Chih-Hui Ho, Kuan-Chuan Peng, and Nuno Vasconcelos. Long-tailed anomaly detection with learnable class names. In Proc. CVPR, 2024. 1, 2, 3, 5, 6, 7, 8, 4, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21

  2. [1]

    CableInspect-AD: An expert- annotated anomaly detection dataset

    Akshatha Arodi, Margaux Luck, Jean-Luc Bedwani, Aldo Zaimi, Ge Li, Nicolas Pouliot, Julien Beaudry, and Ga´etan Marceau Caron. CableInspect-AD: An expert- annotated anomaly detection dataset. In Proc. NeurIPS,

  3. [2]

    Dual-path frequency discriminators for few-shot anomaly detection.Knowledge- Based Systems, 2024

    Yuhu Bai, Jiangning Zhang, Zhaofeng Chen, Yuhang Dong, Yunkang Cao, and Guanzhong Tian. Dual-path frequency discriminators for few-shot anomaly detection.Knowledge- Based Systems, 2024. 8

  4. [3]

    The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification

    Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 5

  5. [4]

    BMAD: Benchmarks for med- ical anomaly detection

    Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaox- iang Zhang, and Xingyu Li. BMAD: Benchmarks for med- ical anomaly detection. In Proc. CVPR, 2024. 5, 1

  6. [5]

    MVTec AD–a comprehensive real-world dataset for unsupervised anomaly detection

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. MVTec AD–a comprehensive real-world dataset for unsupervised anomaly detection. In Proc. CVPR, 2019. 1, 5, 3, 4

  7. [6]

    The liver tumor segmentation benchmark (lits)

    Patrick Bilic, Patrick Christ, Hongwei Bran Li, Eugene V orontsov, Avi Ben-Cohen, Georgios Kaissis, Adi Szeskin, Colin Jacobs, Gabriel Efrain Humpire Mamani, Gabriel Chartrand, et al. The liver tumor segmentation benchmark (lits). Medical Image Analysis, 2023. 5

  8. [7]

    Segment any anomaly without training via hybrid prompt regularization

    Yunkang Cao, Xiaohao Xu, Chen Sun, Yuqi Cheng, Zong- wei Du, Liang Gao, and Weiming Shen. Segment any anomaly without training via hybrid prompt regularization. arXiv preprint arXiv:2305.10724, 2023. 1, 2, 8

Show all 149 references
  1. [8]

    A survey on visual anomaly detection: Challenge, approach, and prospect

    Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1

  2. [9]

    AdaCLIP: Adapt- ing CLIP with hybrid learnable prompts for zero-shot anomaly detection

    Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. AdaCLIP: Adapt- ing CLIP with hybrid learnable prompts for zero-shot anomaly detection. In Proc. ECCV, 2024. 1, 2, 8

  3. [10]

    Human-free au- tomated prompting for vision-language anomaly detection: Prompt optimization with meta-guiding prompt scheme

    Pi-Wei Chen, Jerry Chun-Wei Lin, Jia Ji, Feng-Hao Yeh, Zih-Ching Chen, and Chao-Chun Chen. Human-free au- tomated prompting for vision-language anomaly detection: Prompt optimization with meta-guiding prompt scheme. arXiv preprint arXiv:2406.18197, 2024. 1, 2, 8

  4. [11]

    A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization

    Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. In Proc. ECCV, 2024. 8

  5. [12]

    Sub-image anomaly detec- tion with deep pyramid correspondences

    Niv Cohen and Yedid Hoshen. Sub-image anomaly detec- tion with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 8

  6. [13]

    Microsoft Copilot, 2023

    Microsoft Corporation. Microsoft Copilot, 2023. 4, 2

  7. [14]

    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. InProc. ICPR, 2021. 8

  8. [15]

    Anomaly detection via reverse distillation from one-class embedding

    Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proc. CVPR, 2022. 8

  9. [16]

    Continual learning for anomaly detection in surveillance videos

    Keval Doshi and Yasin Yilmaz. Continual learning for anomaly detection in surveillance videos. In Proc. CVPR Workshop, 2020. 1

  10. [17]

    Transformers: Align model docu- mentation

    Hugging Face. Transformers: Align model docu- mentation. https : / / huggingface . co / docs / transformers/model_doc/align. 2

  11. [18]

    ChangeChip: A reference-based unsupervised change de- tection for PCB defect detection

    Yehonatan Fridman, Matan Rusanovsky, and Gal Oren. ChangeChip: A reference-based unsupervised change de- tection for PCB defect detection. In Proc. PAINE, 2021. 1

  12. [19]

    TransFusion–a transparency-based diffusion model for anomaly detection

    Matic Fu ˇcka, Vitjan Zavrtanik, and Danijel Sko ˇcaj. TransFusion–a transparency-based diffusion model for anomaly detection. In Proc. ECCV, 2024. 8

  13. [20]

    Leveraging vector-quantized variational autoencoder inner metrics for anomaly detection

    Hugo Gangloff, Minh-Tan Pham, Luc Courtrai, and S´ebastien Lef`evre. Leveraging vector-quantized variational autoencoder inner metrics for anomaly detection. In Proc. ICPR, 2022. 1, 5

  14. [21]

    Learning to detect multi-class anomalies with just one normal image prompt

    Bin-Bin Gao. Learning to detect multi-class anomalies with just one normal image prompt. In Proc. ECCV, 2024. 5, 8

  15. [22]

    Towards total online unsupervised anomaly detection and localization in industrial vision

    Han Gao, Huiyuan Luo, Fei Shen, and Zhengtao Zhang. Towards total online unsupervised anomaly detection and localization in industrial vision. arXiv preprint arXiv:2305.15652, 2023. 1, 2, 8

  16. [23]

    Real-time evaluation in online continual learning: A new hope

    Yasir Ghunaim, Adel Bibi, Kumail Alhamoud, Motasem Alfarra, Hasan Abed Al Kader Hammoud, Ameya Prabhu, Philip HS Torr, and Bernard Ghanem. Real-time evaluation in online continual learning: A new hope. In Proc. CVPR,

  17. [24]

    Filo: Zero-shot anomaly detection by fine-grained description and high- quality localization

    Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Hao Li, Ming Tang, and Jinqiao Wang. Filo: Zero-shot anomaly detection by fine-grained description and high- quality localization. In Proc. ACM MM, 2024. 8

  18. [25]

    AnomalyGPT: Detecting in- dustrial anomalies using large vision-language models

    Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. AnomalyGPT: Detecting in- dustrial anomalies using large vision-language models. In Proc. AAAI, 2024. 2, 6, 8, 5, 7, 9, 10, 11, 12, 13, 14

  19. [26]

    Few-shot anomaly-driven generation for anomaly classification and segmentation

    Guan Gui, Bin-Bin Gao, Jun Liu, Chengjie Wang, and Yunsheng Wu. Few-shot anomaly-driven generation for anomaly classification and segmentation. In Proc. ECCV,

  20. [27]

    Absolute-unified multi-class anomaly detection via class-agnostic distribution alignment

    Jia Guo, Haonan Han, Shuai Lu, Weihang Zhang, and Huiqi Li. Absolute-unified multi-class anomaly detection via class-agnostic distribution alignment. arXiv preprint arXiv:2404.00724, 2024. 8

  21. [28]

    Br35H: Brain tumor detection 2020,

    Ahmed Hamada. Br35H: Brain tumor detection 2020,

  22. [29]

    OneLLM: One framework to align all modali- ties with language

    Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xi- angyu Yue. OneLLM: One framework to align all modali- ties with language. In Proc. CVPR, 2024. 8

  23. [30]

    MambaAD: Exploring state space models for multi-class unsupervised anomaly detec- tion

    Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. MambaAD: Exploring state space models for multi-class unsupervised anomaly detec- tion. In Proc. NeurIPS, 2024. 2, 8

  24. [31]

    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. A diffusion-based framework for multi-class anomaly detection. In Proc. AAAI, 2024. 8

  25. [32]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, 2016. 4

  26. [33]

    Learning unified reference rep- resentation for unsupervised multi-class anomaly detection

    Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qian- gang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference rep- resentation for unsupervised multi-class anomaly detection. In Proc. ECCV, 2024. 8

  27. [35]

    Automated seg- mentation of macular edema in oct using deep neural net- works

    Junjie Hu, Yuanyuan Chen, and Zhang Yi. Automated seg- mentation of macular edema in oct using deep neural net- works. Medical image analysis, 2019. 5

  28. [36]

    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. In Proc. ECCV, 2022. 6, 5, 7, 8, 9, 10, 11, 12, 13, 14

  29. [37]

    Adapting visual- language models for generalizable anomaly detection in medical images

    Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xinchao Wang, and Yanfeng Wang. Adapting visual- language models for generalizable anomaly detection in medical images. In Proc. CVPR, 2024. 1, 2, 8

  30. [38]

    ReCon- Patch: Contrastive patch representation learning for indus- trial anomaly detection

    Jeeho Hyun, Sangyun Kim, Giyoung Jeon, Seung Hwan Kim, Kyunghoon Bae, and Byung Jun Kang. ReCon- Patch: Contrastive patch representation learning for indus- trial anomaly detection. In Proc. WACV, 2024. 8

  31. [39]

    Towards open-world object-based anomaly detection via self-supervised outlier synthesis

    Brian KS Isaac-Medina, Yona Falinie A Gaus, Neelan- jan Bhowmik, and Toby P Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In Proc. ECCV, 2024. 8

  32. [40]

    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. InProc. CVPR, 2023. 1, 2, 5, 8

  33. [41]

    Scaling up visual and vision-language representation learning with noisy text supervision

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. InProc. ICML, 2021. 8, 2, 4

  34. [42]

    Brain tumor detec- tion using MRI images

    Pranita Balaji Kanade and PP Gumaste. Brain tumor detec- tion using MRI images. Brain, 2015. 1

  35. [43]

    Head CT - hemorrhage, 2018

    Felipe Campos Kitamura. Head CT - hemorrhage, 2018. 1

  36. [44]

    Online continual learning on class incremental blurry task configuration with anytime inference

    Hyunseo Koh, Dahyun Kim, Jung-Woo Ha, and Jonghyun Choi. Online continual learning on class incremental blurry task configuration with anytime inference. In Proc. ICLR,

  37. [45]

    Miccai multi-atlas la- beling beyond the cranial vault–workshop and challenge

    Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas la- beling beyond the cranial vault–workshop and challenge. In Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, 2015. 5

  38. [46]

    Continuous memory representation for anomaly detection

    Joo Chan Lee, Taejune Kim, Eunbyung Park, Simon S Woo, and Jong Hwan Ko. Continuous memory representation for anomaly detection. In Proc. ECCV, 2024. 8

  39. [47]

    AD3: Introducing a score for anomaly detec- tion dataset difficulty assessment using VIADUCT dataset

    Jan Lehr, Jan Philipps, Alik Sargsyan, Martin Pape, and J¨org Kr¨uger. AD3: Introducing a score for anomaly detec- tion dataset difficulty assessment using VIADUCT dataset. In Proc. ECCV, 2024. 1

  40. [48]

    CutPaste: Self-supervised learning for anomaly de- tection and localization

    Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. CutPaste: Self-supervised learning for anomaly de- tection and localization. In Proc. CVPR, 2021. 8

  41. [49]

    ClipSAM: CLIP and SAM collabora- tion for zero-shot anomaly segmentation

    Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. ClipSAM: CLIP and SAM collabora- tion for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 2, 8

  42. [50]

    MuSc: Zero-shot industrial anomaly classification and segmenta- tion with mutual scoring of the unlabeled images

    Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. MuSc: Zero-shot industrial anomaly classification and segmenta- tion with mutual scoring of the unlabeled images. In Proc. ICLR, 2024. 8

  43. [51]

    PromptAD: Learning prompts with only normal samples for few-shot anomaly 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 anomaly detection. In Proc. CVPR, 2024. 1, 2, 5, 8

  44. [52]

    FADE: Few-shot/zero-shot anomaly detection engine using large vision-language model

    Yuanwei Li, Elizaveta Ivanova, and Martins Bruveris. FADE: Few-shot/zero-shot anomaly detection engine using large vision-language model. In Proc. BMVC, 2024. 1, 2, 8

  45. [53]

    COFT-AD: Contrastive fine-tuning for few-shot anomaly detection

    Jingyi Liao, Xun Xu, Manh Cuong Nguyen, Adam Goodge, and Chuan Sheng Foo. COFT-AD: Contrastive fine-tuning for few-shot anomaly detection. IEEE TIP, 2024. 8

  46. [54]

    Learning diffusion models for multi-view anomaly detection

    Chieh Liu, Yu-Min Chu, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning diffusion models for multi-view anomaly detection. In Proc. ECCV, 2024. 8

  47. [55]

    Unsupervised continual anomaly detection with contrastively-learned prompt

    Jiaqi Liu, Kai Wu, Qiang Nie, Ying Chen, Bin-Bin Gao, Yong Liu, Jinbao Wang, Chengjie Wang, and Feng Zheng. Unsupervised continual anomaly detection with contrastively-learned prompt. In Proc. AAAI, 2024. 1, 8

  48. [56]

    Heterogeneity-aware recurrent neu- ral network for hyperspectral and multispectral image fu- sion

    Ruiying Lu, Bo Chen, Jianqiao Sun, Wenchao Chen, Penghui Wang, Yuanwei Chen, Hongwei Liu, and Pramod K Varshney. Heterogeneity-aware recurrent neu- ral network for hyperspectral and multispectral image fu- sion. IEEE Journal of Selected Topics in Signal Processing,

  49. [57]

    Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection

    Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. In Proc. NeurIPS, 2023. 1, 2, 4, 5, 6, 8, 7, 9, 10, 11, 12, 13, 14

  50. [58]

    Review of wafer surface defect detection methods

    Jianhong Ma, Tao Zhang, Cong Yang, Yangjie Cao, Lipeng Xie, Hui Tian, and Xuexiang Li. Review of wafer surface defect detection methods. Electronics, 2023. 1

  51. [59]

    Anomaly de- tection through latent space restoration using vector quan- tized variational autoencoders

    Sergio Naval Marimont and Giacomo Tarroni. Anomaly de- tection through latent space restoration using vector quan- tized variational autoencoders. In Proc. ISBI, 2021. 1, 5, 8

  52. [60]

    Mixture of ex- perts: a literature survey

    Saeed Masoudnia and Reza Ebrahimpour. Mixture of ex- perts: a literature survey. Artificial Intelligence Review ,

  53. [61]

    Unsu- pervised, online and on-the-fly anomaly detection for non- stationary image distributions

    Declan GD McIntosh and Alexandra Branzan Albu. Unsu- pervised, online and on-the-fly anomaly detection for non- stationary image distributions. In Proc. ECCV, 2024. 1, 2, 8

  54. [62]

    MoEAD: A parameter- efficient model for multi-class anomaly detection

    Shiyuan Meng, Wenchao Meng, Qihang Zhou, Shizhong Li, Weiye Hou, and Shibo He. MoEAD: A parameter- efficient model for multi-class anomaly detection. In Proc. ECCV, 2024. 2, 6, 8, 5, 7, 9, 10, 11, 12, 13, 14

  55. [63]

    RGI: Robust GAN- inversion for mask-free image inpainting and unsupervised pixel-wise anomaly detection

    Shancong Mou, Xiaoyi Gu, Meng Cao, Haoping Bai, Ping Huang, Jiulong Shan, and Jianjun Shi. RGI: Robust GAN- inversion for mask-free image inpainting and unsupervised pixel-wise anomaly detection. In Proc. ICLR, 2023. 8

  56. [64]

    ChatGPT, 2023

    OpenAI. ChatGPT, 2023. 4, 8, 2

  57. [65]

    Deep learning for anomaly detection: A review

    Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM computing surveys (CSUR), 2021. 1

  58. [66]

    Revisiting deep feature reconstruction for logical and structural industrial anomaly detection

    Sukanya Patra and Souhaib Ben Taieb. Revisiting deep feature reconstruction for logical and structural industrial anomaly detection. TMLR, 2024. 8

  59. [67]

    VCP-CLIP: A visual context prompting model for zero-shot anomaly seg- mentation

    Zhen Qu, Xian Tao, Mukesh Prasad, Fei Shen, Zhengtao Zhang, Xinyi Gong, and Guiguang Ding. VCP-CLIP: A visual context prompting model for zero-shot anomaly seg- mentation. In Proc. ECCV, 2024. 1, 2, 8

  60. [68]

    Learn- ing transferable visual models from natural language super- vision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learn- ing transferable visual models from natural language super- vision. In Proc. ICML, 2021. 8

  61. [69]

    DELTA: Decoupling long-tailed online continual learning

    Siddeshwar Raghavan, Jiangpeng He, and Fengqing Zhu. DELTA: Decoupling long-tailed online continual learning. In Proc. CVPR, 2024. 8

  62. [70]

    Variational infer- ence with normalizing flows

    Danilo Rezende and Shakir Mohamed. Variational infer- ence with normalizing flows. In Proc. ICML, 2015. 8

  63. [71]

    Scaling vision with sparse mix- ture of experts

    Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, Andr´e Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mix- ture of experts. In Proc. NeurIPS, 2021. 8

  64. [72]

    Towards to- tal recall in industrial anomaly detection

    Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Sch¨olkopf, Thomas Brox, and Peter Gehler. Towards to- tal recall in industrial anomaly detection. In Proc. CVPR,

  65. [73]

    Same same but differnet: Semi-supervised defect detection with normalizing flows

    Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proc. WACV, 2021. 8

  66. [74]

    Tire defect detection model using ma- chine learning

    Radhwan AA Saleh, Mehmet Zeki Konyar, Kaplan Kaplan, and H Metin Ertunc ¸. Tire defect detection model using ma- chine learning. In Proc. eSmarTA, 2022. 1

  67. [75]

    Multiresolution knowledge distillation for anomaly detection

    Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Ra- biee. Multiresolution knowledge distillation for anomaly detection. In Proc. CVPR, 2021. 8, 6, 7

  68. [76]

    Dissolving is amplifying: Towards fine-grained anomaly detection

    Jian Shi, Pengyi Zhang, Ni Zhang, Hakim Ghazzai, and Pe- ter Wonka. Dissolving is amplifying: Towards fine-grained anomaly detection. In Proc. ECCV, 2024. 8

  69. [77]

    GeneralAD: Anomaly detection across domains by attending to distorted features

    Luc PJ Str ¨ater, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. GeneralAD: Anomaly detection across domains by attending to distorted features. In Proc. ECCV, 2024. 8

  70. [78]

    EfficientNet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, 2019. 2

  71. [79]

    An incremental unified framework for small defect inspection

    Jiaqi Tang, Hao Lu, Xiaogang Xu, Ruizheng Wu, Sixing Hu, Tong Zhang, Tsz Wa Cheng, Ming Ge, Ying-Cong Chen, and Fugee Tsung. An incremental unified framework for small defect inspection. In Proc. ECCV, 2024. 1, 8

  72. [80]

    Revisiting reverse distillation for anomaly detection

    Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan Duong, Chanh D Tr Nguyen, and Steven QH Truong. Revisiting reverse distillation for anomaly detection. In Proc. CVPR, 2023. 8

  73. [81]

    LLaMA: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth ´ee Lacroix, Bap- tiste Rozi `ere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 8

  74. [82]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In Proc. NeurIPS, 2017. 5, 8

  75. [83]

    A comprehensive survey of continual learning: theory, method and application

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. IEEE TPAMI, 2024. 8

  76. [84]

    Few-shot online anomaly detection and segmentation

    Shenxing Wei, Xing Wei, Zhiheng Ma, Shaochen Zhang Songlin Dong, and Yihong Gong. Few-shot online anomaly detection and segmentation. Knowledge-Based Systems ,

  77. [85]

    Weakly supervised learning for industrial optical inspection

    Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM sym- posium in, 2007. 5, 3, 4

  78. [86]

    Defect spectrum: A granular look of large-scale defect datasets with rich seman- tics

    Shuai Yang, Zhifei Chen, Pengguang Chen, Xi Fang, Yixun Liang, Shu Liu, and Yingcong Chen. Defect spectrum: A granular look of large-scale defect datasets with rich seman- tics. In Proc. ECCV, 2024. 1

  79. [87]

    GLAD: Towards better reconstruction with global and local adaptive diffu- sion models for unsupervised anomaly detection

    Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. GLAD: Towards better reconstruction with global and local adaptive diffu- sion models for unsupervised anomaly detection. In Proc. ECCV, 2024. 8

  80. [88]

    Hierarchical gaussian mixture normal- izing flow modeling for unified anomaly detection

    Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normal- izing flow modeling for unified anomaly detection. InProc. ECCV, 2024

  81. [89]

    A unified model for multi-class anomaly detection

    Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Proc. NeurIPS, 2022. 5, 6, 8, 7, 9, 10, 11, 12, 13, 14

  82. [90]

    FastFlow: Unsupervised anomaly detection and localization via 2D normalizing flows

    Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. FastFlow: Unsupervised anomaly detection and localization via 2D normalizing flows. arXiv preprint arXiv:2111.07677, 2021. 8

  83. [91]

    DRAEM-a discriminatively trained reconstruction embed- ding for surface anomaly detection

    Vitjan Zavrtanik, Matej Kristan, and Danijel Sko ˇcaj. DRAEM-a discriminatively trained reconstruction embed- ding for surface anomaly detection. In Proc. ICCV, 2021. 8, 6, 7

  84. [92]

    A systematic review on long- tailed learning

    Chongsheng Zhang, George Almpanidis, Gaojuan Fan, Binquan Deng, Yanbo Zhang, Ji Liu, Aouaidjia Kamel, Paolo Soda, and Jo˜ao Gama. A systematic review on long- tailed learning. arXiv preprint arXiv:2408.00483, 2024. 1

  85. [93]

    Defect-GAN: High-fidelity defect synthesis for automated defect inspection

    Gongjie Zhang, Kaiwen Cui, Tzu-Yi Hung, and Shijian Lu. Defect-GAN: High-fidelity defect synthesis for automated defect inspection. In Proc. WACV, 2021. 8

  86. [94]

    Exploring grounding poten- tial of VQA-oriented GPT-4V for zero-shot anomaly detec- tion

    Jiangning Zhang, Xuhai Chen, Zhucun Xue, Yabiao Wang, Chengjie Wang, and Yong Liu. Exploring grounding poten- tial of VQA-oriented GPT-4V for zero-shot anomaly detec- tion. arXiv preprint arXiv:2311.02612, 2023. 2, 8

  87. [95]

    Exploring plain ViT reconstruction for multi- class unsupervised anomaly detection

    Jiangning Zhang, Xuhai Chen, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, Ming-Hsuan Yang, and Dacheng Tao. Exploring plain ViT reconstruction for multi- class unsupervised anomaly detection. In Proc. IJCAI Workshop on Anomaly Detection with Foundation Models,

  88. [96]

    ADer: A comprehensive benchmark for multi-class visual anomaly detection

    Jiangning Zhang, Haoyang He, Zhenye Gan, Qingdong He, Yuxuan Cai, Zhucun Xue, Yabiao Wang, Chengjie Wang, Lei Xie, and Yong Liu. ADer: A comprehensive benchmark for multi-class visual anomaly detection. arXiv preprint arXiv:2406.03262, 2024. 1, 4

  89. [97]

    DeSTSeg: Segmentation guided denois- ing student-teacher for anomaly detection

    Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. DeSTSeg: Segmentation guided denois- ing student-teacher for anomaly detection. In Proc. CVPR,

  90. [98]

    Meta- Transformer: A unified framework for multimodal learn- ing

    Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hong- sheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. Meta- Transformer: A unified framework for multimodal learn- ing. arXiv preprint arXiv:2307.10802, 2023. 8

  91. [99]

    OmniAL: A unified cnn framework for unsu- pervised anomaly localization

    Ying Zhao. OmniAL: A unified cnn framework for unsu- pervised anomaly localization. In Proc. CVPR, 2023. 8

  92. [100]

    AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection

    Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jim- ing Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In Proc. ICLR, 2024. 1, 2, 8

  93. [101]

    Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts

    Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proc. CVPR, 2024. 8

  94. [102]

    Fine-grained abnormality prompt learning for zero- shot anomaly detection

    Jiawen Zhu, Yew-Soon Ong, Chunhua Shen, and Guansong Pang. Fine-grained abnormality prompt learning for zero- shot anomaly detection. arXiv preprint arXiv:2410.10289,

  95. [103]

    hard- est

    Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self- supervised pre-training for anomaly detection and segmen- tation. In Proc. ECCV, 2022. 1, 5, 3, 4 Appendix The appendix is organized as follows: • In Sec. A1, we provide additio...

  96. [104]

    Out of 100 [bc] which looks generally the same, one appears to have a broken region. List 5 examples

    As a comparison, HVQ [57] samples 512 codes of di- mension 256 for their K codebooks. Overall, we require far fewer codes than since 16 × 10 × 640 ≤ 512K × 256 for all K ≥ 1. For bY (see Eq. 4), we set α = 0.3. In AAA (see Alg. 1), we set γ = 0.3, β = 5, τ = 0.2, and T ( bY ) ...

  97. [105]

    carpet, hazelnut, metal nut

  98. [106]

    bottle, screw, capsule

  99. [108]

    We organize the classes in each session as the following list, i.e.D5-TF in reverse order

    pill, transistor, grid On the other hand, inD5-TF, we divide theDO into 5 dis- joint sessions and we put the Ctail in the first few sessions. We organize the classes in each session as the following list, i.e.D5-TF in reverse order

  100. [109]

    pill, transistor, grid

  101. [110]

    zipper, cable, toothbrush

  102. [111]

    screw, capsule, bottle

  103. [112]

    We organize the classes in each session as the following list

    carpet, hazelnut, metal nut In D5-M, we divide the DO into 5 disjoint sessions and we put both Chead and Ctail in each session. We organize the classes in each session as the following list

  104. [113]

    hazelnut, transistor, capsule,

  105. [114]

    metal nut, wood, screw

  106. [115]

    leather, pill, toothbrush

  107. [116]

    Let the number of remain- ing Chead images at step t, i.e

    bottle, tile, zipper In B-HF, for each i ∈ 1, · · ·, NO where N O = |DO|, we use the following operation to determine where fXi should be sampled from Chead or Ctail. Let the number of remain- ing Chead images at step t, i.e. DO[t :] be N head t and the one for Ctail be N tail...

  108. [119]

    candle, capsules, cashew In D5-TF, we organize the classes in each session as the following list

  109. [120]

    candle, capsules, cashew

  110. [121]

    chewinggum, fryum, pipe fryum

  111. [122]

    macaroni1, macaroni2

  112. [123]

    pcb3, pcb4 In D5-M, we organize the classes in each session as the fol- lowing list

  113. [124]

    capsules, macaroni2, pcb1

  114. [125]

    DAGM DAGM is a synthetic anomaly detection dataset, containing 10 classes

    fryum, pipe fryum, pcb4 A2.3. DAGM DAGM is a synthetic anomaly detection dataset, containing 10 classes. We list out theC in the 3rd row of Tab. A1. Each class has a different synthetic pattern. Here we use the same design logic for the configurations in MVTec. In D5-HF, we or...

  115. [126]

    Class1, Class3 In D5-TF, we organize the classes in each session as the following list

  116. [127]

    Class9, Class10 In D5-M, we organize the classes in each session as the following list

  117. [128]

    Offline experiments We provide additional analysis of our approach

    Class10, Class6 A3. Offline experiments We provide additional analysis of our approach. We then report the performances on Chead, on Ctail, and on each c ∈ C in the following tables under various datasets and long- tailed imbalance [34] settings. A3.1. Additional analysis Alte...

  118. [129]

    A11 and Tab

    exp100: Tab. A11 and Tab. A12

  119. [130]

    A13 and Tab

    exp200: Tab. A13 and Tab. A14

  120. [131]

    A15 and Tab

    step100: Tab. A15 and Tab. A16

  121. [132]

    A17 and Tab

    step200: Tab. A17 and Tab. A18. • VisA

  122. [133]

    A19 and Tab

    exp100: Tab. A19 and Tab. A20

  123. [134]

    A21 and Tab

    exp200: Tab. A21 and Tab. A22

  124. [135]

    A23 and Tab

    exp500: Tab. A23 and Tab. A24

  125. [136]

    A25 and Tab

    step100: Tab. A25 and Tab. A26

  126. [137]

    A27 and Tab

    step200: Tab. A27 and Tab. A28

  127. [138]

    A29 and Tab

    step500: Tab. A29 and Tab. A30. • DAGM

  128. [139]

    A31 and Tab

    exp50: Tab. A31 and Tab. A32

  129. [140]

    A33 and Tab

    exp100: Tab. A33 and Tab. A34

  130. [141]

    A35 and Tab

    exp200: Tab. A35 and Tab. A36

  131. [142]

    A37 and Tab

    exp500: Tab. A37 and Tab. A38

  132. [143]

    A39 and Tab

    reverse exp200: Tab. A39 and Tab. A40

  133. [144]

    A41 and Tab

    step50: Tab. A41 and Tab. A42

  134. [145]

    A43 and Tab

    step100: Tab. A43 and Tab. A44

  135. [146]

    A45 and Tab

    step200: Tab. A45 and Tab. A46

  136. [147]

    A47 and Tab

    step500: Tab. A47 and Tab. A48

  137. [148]

    A49 and Tab

    reverse step200: Tab. A49 and Tab. A50.. We note that in reverse exp200 and reverse step200, LTAD switches the Chead and Ctail. We observe a consistent im- provement of LTOAD in comparison to all baselines on most configurations. Notably, LTOAD excels on the more challenging C...

  138. [149]

    step100: Fig. A7. Method CA C C head Ctail brain liver retinalDet. Seg. Det. Seg. Det. Seg. Det. Seg. Det. Seg. Det. Seg. HVQ [57]✗ 65.21 93.84 83.15 97.03 56.23 92.25 83.15 97.03 45.7295.4266.74 89.07MoEAD [62]✓60.16 91.0986.74 97.5046.88 87.8886.74 97.5043.60 94.82 50.15 80....

  139. [150]

    step200: Fig. A8. • VisA

  140. [151]

    step200: Fig. A12. • DAGM

  141. [152]

    step200: Fig. A16. Apart from the synthetic DAGM dataset where the offline performance is saturated, we observe that LTOAD improves the performances in most cases, especially under the more challenging offline long-tailed settings, i.e. step200. Method CA C Chead Ctailhazelnut...

Pith tools

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