Pith. sign in

REVIEW 4 major objections 4 minor 43 references

Generative Model-Based Feature Attention Module for Video Action Analysis

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that few-shot anomaly segmentation—finding defective regions in images of object classes never seen in training—can be reduced to a dictionary lookup: build an index from a few normal reference images, and mark any query r

desk verdict The submitted record and the manuscript are two different papers; the actual PDF (DictAS) is a strong FSAS submission with a real SOTA claim, but the training protocol leaves an instance-matching shortcut that a referee should probe. read the letter →

arxiv 2508.13565 v1 pith:LHHJOXGT submitted 2025-08-19 cs.CV

Guiqin Wang , Peng Zhao , Cong Zhao , Jing Huang , Siyan Guo , Shusen Yang This is my paper
classification cs.CV
keywords few-shotanomalysegmentationclass-generalizabledictionarylookupCLIPself-supervisedlearningdetectionsparsemaxindustrialdefect
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

This paper claims that few-shot anomaly segmentation—finding defective regions in images of object classes never seen in training—can be reduced to a dictionary lookup: build an index from a few normal reference images, and mark any query region that cannot be retrieved from that index as an anomaly. Prior few-shot methods, it argues, rely on prior knowledge from real abnormal samples and so generalize poorly to novel classes; a self-supervised retrieval skill transfers instead. The paper builds DictAS on a frozen CLIP encoder, trains it on unlabeled images with synthetic defects, and reports state-of-the-art pixel-level segmentation on five industrial and two medical datasets, including gains over a class-dependent method that fine-tunes per class. If correct, this points toward inspection systems that switch product lines or medical tasks without retraining.

What carries the argument

The load-bearing mechanism is the dictionary lookup pipeline: a Value Generator preserves reference detail with a residual connection, a Key Generator indexes it, and a Query Generator transforms the query; matching uses sparsemax, a sparse attention weight that keeps only the most relevant dictionary entries rather than spreading mass densely. The dictionary is trained self-supervised, with DRAEM's Perlin-noise anomaly synthesis generating query images and pseudo-labels for the two discrimination regularizers. This combination is what lets a single model transfer the retrieval operation—rather than memorized normal/abnormal patterns—to classes it has never seen.

What would settle it

Take a class with subtle, high-variance textures (for example, the foam category of MVTec3D or fine scratches on brushed metal) and inspect the per-patch retrieval-distance histograms for normal versus anomalous regions: the claim predicts near-zero distance for normals and clearly separated, larger distances for anomalies. If the two distributions overlap substantially, or if anomalous patches are retrievable as combinations of dictionary entries, the cosine-distance anomaly map cannot separate them. A cleaner controlled test is to build reference dictionaries from one imaging domain (polishe

Watch

Extended reading notes

Core claim

The central claim is that a model can learn a class-agnostic dictionary lookup skill: given a query image and a handful of normal reference images of an unseen class, a patch is normal if its feature can be matched back into a dictionary of reference features and anomalous if it cannot. The paper implements this with three learned generators (Key, Value, Query) built on self-attention blocks, a sparse lookup strategy (sparsemax) that fuses only the most relevant dictionary entries, and a query loss that uses the query feature itself as a pseudo-label for normal regions. Two regularizers—a Contrastive Query Constraint and a Text Alignment Constraint—push anomalous features out of the retrieva

Load-bearing premise

The method assumes that every normal patch of a query image is retrievable from a dictionary built from a few normal reference images of the same class, while every anomalous patch is not—so the whole segmentation collapses if anomalies can be reassembled from combinations of normal features, or if normal regions fail to match the reference dictionary.

Editorial extensions

If this is right

  • One unified model, trained without any pixel-level annotations, segments anomalies in unseen industrial and medical classes from a few normal reference images.
  • Performance keeps improving as the number of reference images grows from 1 to 16 shots, whereas competing methods plateau or decline.
  • Auxiliary training data can come from any domain—industrial, medical, or natural scenes—so no annotated corpus is required to build the model.
  • The method is the fastest CLIP-based few-shot anomaly segmenter tested, at 73.5 ms per image on MVTecAD under the 4-shot setting.
  • It outperforms a per-class fine-tuned baseline (PromptAD), which is usually the stronger setting, without any per-class fine-tuning.

Reading between the lines

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

  • If the retrieval-based framing holds, the same dictionary-lookup objective could extend to other open-set vision problems—novel-class detection, change detection, or visual inspection under distribution shift—where 'normal' is defined by a few exemplars rather than a fixed training set.
  • The reliance on DRAEM's synthetic Perlin-noise anomalies is the most likely transfer bottleneck; a natural testable extension is training on harder synthetic anomalies (object-part defects, text-guided edits) and measuring the gain on the weakest reported categories (foam, PCBs).
  • Reference-free, true zero-shot anomaly detection is the obvious next frontier; the paper states that the dictionary requirement currently blocks it.
  • Editorial note: the document's front-matter abstract and author list describe a different manuscript on video action analysis, while the substantive body is the DictAS paper; the supplied arXiv identifier (2508.13565) also disagrees with the paper's own header (2508.13560v2), so readers should verify the record before citing.
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 DictAS, a self-supervised framework for class-generalizable few-shot anomaly segmentation (FSAS). It builds a dictionary from a few normal reference images using learned Key/Value generators, retrieves query patch features via sparsemax-based sparse lookup, and computes an anomaly map from the cosine distance between query and retrieved features. Training is self-supervised on raw images without pixel annotations: reference images are transformed versions of the same image, query images are the same image with DRAEM-synthesized anomalies, and the losses are a query reconstruction loss plus two regularizers (contrastive query constraint and text alignment constraint). Experiments on seven industrial and medical datasets under 1-16 shots report state-of-the-art pixel-level AUROC/PRO/AP and favorable inference efficiency. The paper includes component ablations, lookup-strategy ablations, backbone/resolution ablations, auxiliary-dataset ablations, and per-category results. The central claim is that a single model trained on seen classes transfers to unseen classes using only a few normal reference images as visual prompts.

Significance. If the central claim is sound, the contribution is significant: it removes the need for target-class fine-tuning or real anomaly samples, requires no pixel-level annotations, and demonstrates strong results across diverse industrial and medical domains. The paper is also commendable for its breadth of experiments (seven datasets, multiple shot settings, per-category statistics, efficiency measurements, and an ablation of auxiliary training data) and for releasing code. The self-supervised formulation and the dictionary-lookup perspective are interesting and potentially useful for practical FSAS. However, the claimed class-generalizable dictionary lookup capability is not established by the current training protocol, because training is performed with query/reference pairs derived from the same image. The significance of the reported SOTA results is further tempered by the selection of regularization weights on the target test benchmarks. These issues are load-bearing for the main contribution.

major comments (4)
  1. [§3.3, Eq. (7); §3.1 and Algorithm 2 (Appendix A.2)] The training protocol builds the reference image and the query image from the same raw image X: the reference is a transformed version of X, and the query is X with synthesized anomalies. For every normal patch in the query, the dictionary contains features from the exact same image (only transformed). Minimizing Lq in Eq. (7) can therefore be solved by instance-level matching, without learning to retrieve across different instances of a class. At inference, however, the dictionary is constructed from k different normal images of an unseen class (§3.5), which is a categorically different retrieval problem. No experiment separates these two cases; e.g., training with a reference image drawn from a different image of the same class, or an ablation comparing same-image vs cross-image reference construction, is missing. Without such evidence, the claim that DictAS learns a 'category-agnostic
  2. [§C.1, Figure 8; Table 1] The regularization weights λ1 and λ2 are ablated on MVTecAD and BTAD under the 4-shot setting, and these same datasets appear in the main results (Table 1). This constitutes tuning on the target test benchmarks. The reported SOTA numbers are therefore partly the result of configuration selection on the test sets. The authors should either tune on a held-out validation split or auxiliary seen classes, or report results across a range of λ values on all datasets to demonstrate that the choice is not dataset-specific. The 'consistently outperforms' claim needs this unbiased evaluation.
  3. [Title/Abstract vs. Full Text] The document's top-level title and abstract describe 'Generative Model-Based Feature Attention Module for Video Action Analysis', with a code URL matching that different project, while the full text is the DictAS paper on few-shot anomaly segmentation. This is not a minor typo; as submitted, the manuscript is internally incoherent. The authors must reconcile the title, abstract, and code link with the actual content before the paper can be considered for publication.
  4. [§B.2, Table 7] The comparison with concurrent methods MetaUAS and ResAD is not controlled: the numbers are quoted from the original papers using different backbones (EfficientNet-b4 and ImageBind-Huge, respectively) and different input resolutions. The text claims DictAS 'performs better' against ResAD despite using fewer backbone parameters, but the comparison is not apples-to-apples. Either re-run these methods under the same backbone/resolution settings or clearly present the comparison as indicative only, not as a fair head-to-head.
minor comments (4)
  1. [§4.1, Table 2] Efficiency numbers are reported with different backbones and resolutions for different methods. Please add a sentence clarifying that time/GPU cost is not directly comparable across rows with different backbones, or report efficiency of all methods with the same backbone where possible.
  2. [§4.2] The text refers to 'the quantitative results of ZSAS'; this should be FSAS. There is also a typo 'Maximun Lookup' in Eq. (5), and several unicode placeholders (e.g., '��') in equations and captions that need to be replaced with proper symbols.
  3. [Algorithm 2 (Appendix A.2)] The formula for Xq in Algorithm 2 appears malformed: 'Xq ← �(MA ⊙ A) + (1− �)(MA ⊙ X) +MA ⊙ X' does not match the standard DRAEM composition (Xq = (1-MA)⊙X + MA⊙A). Please correct the equation and define all variables (τ, ε, A) in the algorithm.
  4. [Appendix E.2, Tables 19-22] The per-category tables are valuable, but the category names are inconsistent (e.g., 'BrasTS' vs 'BraTS', 'bangel' vs 'bagel'). Please standardize dataset and category names throughout the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is a held-out empirical generalization result, and no equation or citation reduces the prediction to its training inputs.

full rationale

DictAS's derivation chain is: define the anomaly score as the cosine distance between query features and retrieved dictionary features (Eq. 11), train the dictionary generators and regularizers on DRAEM-synthesized pseudo-labels via the query loss (Eq. 7) and the contrastive/text constraints (Eqs. 8-9), then evaluate on unseen classes. The training objective directly optimizes the same distance that is later used as the anomaly score, but this is standard self-supervised/supervised training, not circularity: the test-time prediction is a held-out anomaly map, not the training target. The query loss uses F_q itself as a pseudo-label, but the final output is not F_q; it is a distance between F_q and the retrieved dictionary features on unseen images. The same-image construction of query/reference pairs could enable an instance-matching shortcut, but that is a generalization-risk argument, not a demonstration that the reported result is equivalent to its inputs by construction; the paper's evidence is cross-class, cross-dataset evaluation plus ablations. The two regularization weights were selected with reference to MVTecAD/BTAD results, which is a test-set tuning concern, but no parameter fitted on a subset is then renamed as a prediction of a closely related quantity. Self-citations (e.g., refs. [26]-[29]) appear in the bibliography/related work but are not load-bearing: the method builds on CLIP, WinCLIP, APRIL-GAN, DRAEM, and sparsemax, all external or independently published. No uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via a self-citation. The dictionary-lookup framing is a reformulation of nearest-neighbor retrieval with learned adapters, but it is implemented and tested empirically rather than presented as a derivation from its own definition. Appendix D's stated limitation (requiring reference images, no zero-shot) is an acknowledgment of scope, not a hidden admission of circularity.

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

The ledger shows a typical deep-learning empirical paper: no invented physical entities, but several domain assumptions (retrievability of normal patches, non-retrievability of anomalies, transferability of synthetic pseudo-labels) and two regularization weights that are effectively fitted using two of the test benchmarks. The framework re-labels attention-based retrieval as a dictionary but adds real components (generators, sparsemax, two regularizers).

free parameters (4)
  • λ1 (weight of Contrastive Query Constraint) = 0.1
    Chosen after ablation on MVTecAD and BTAD 4-shot (Fig. 8), both of which appear in the main results table; selection happens with the target test sets in the loop.
  • λ2 (weight of Text Alignment Constraint) = 0.1
    Same ablation as λ1.
  • Input resolution = 336x336
    Table 5 shows AP still increases at 420 and 518; 336 is chosen as a hand-set speed/accuracy default.
  • Training schedule (epochs, LR, batch size) = 30, 1e-4, 24
    Hand-set without ablations; standard values.
assumptions (5)
  • domain assumption Normal query patches are always retrievable from the dictionary built from normal reference features.
    Explicitly stated in Section 3.3 as the core assumption behind the query loss; if violated on an unseen class, the anomaly map (Eq. 11) loses meaning.
  • domain assumption Anomalous patches are not retrievable from the normal dictionary.
    Section 3.3 opening sentence; this is the mechanism that produces anomaly maps via cosine distance.
  • ad hoc to paper DRAEM synthetic anomalies provide pseudo-labels that transfer to real anomalies in unseen classes.
    Appendix A.2 and Algorithm 2; the entire self-supervised training uses Perlin-noise synthetic masks as G and yq, and the paper offers no direct evidence of transfer beyond benchmark results.
  • domain assumption Frozen CLIP ViT-L-14 features retain enough discriminative signal for unseen-class anomaly separation after lightweight generators.
    Backbone frozen throughout (Section 4.1, A.4); all adaptation is in the small generators.
  • standard math Sparsemax projection is computed correctly by Algorithm 1 (cumulative-sum threshold).
    Adapted from Martins and Astudillo (2016), cited as [23]; standard optimization result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative Model-Based Feature Attention Module for Video Action Analysis." pith.science (2026). https://pith.science/paper/LHHJOXGT

@misc{pith2026250813565,
  author       = {Pith},
  title        = {Pith review of: Generative Model-Based Feature Attention Module for Video Action Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LHHJOXGT}},
  note         = {Machine review of arXiv:2508.13565}
}
read the original abstract

Video action analysis is a foundational technology within the realm of intelligent video comprehension, particularly concerning its application in Internet of Things(IoT). However, existing methodologies overlook feature semantics in feature extraction and focus on optimizing action proposals, thus these solutions are unsuitable for widespread adoption in high-performance IoT applications due to the limitations in precision, such as autonomous driving, which necessitate robust and scalable intelligent video analytics analysis. To address this issue, we propose a novel generative attention-based model to learn the relation of feature semantics. Specifically, by leveraging the differences of actions' foreground and background, our model simultaneously learns the frame- and segment-dependencies of temporal action feature semantics, which takes advantage of feature semantics in the feature extraction effectively. To evaluate the effectiveness of our model, we conduct extensive experiments on two benchmark video task, action recognition and action detection. In the context of action detection tasks, we substantiate the superiority of our approach through comprehensive validation on widely recognized datasets. Moreover, we extend the validation of the effectiveness of our proposed method to a broader task, video action recognition. Our code is available at https://github.com/Generative-Feature-Model/GAF.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 37 canonical work pages

  1. [1]

    Medical image analysis using convolu- tional neural networks: a review.Journal of medical systems, 42:1–13, 2018

    Syed Muhammad Anwar, Muhammad Majid, Adnan Qayyum, Muhammad Awais, Majdi Alnowami, and Muham- mad Khurram Khan. Medical image analysis using convolu- tional neural networks: a review.Journal of medical systems, 42:1–13, 2018. 1

    work page 2018

  2. [2]

    Fewsome: One-class few shot anomaly detection with siamese networks

    Niamh Belton, Misgina Tsighe Hagos, Aonghus Lawlor, and Kathleen M Curran. Fewsome: One-class few shot anomaly detection with siamese networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 2977–2986, 2023. 1, 2

    work page 2023

  3. [3]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592–9600, 2019. 6, 7, 11, 14, 16, 19

    work page 2019

  4. [4]

    The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization

    Paul Bergmann, Xin Jin, David Sattlegger, and Carsten Ste- ger. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. arXiv preprint arXiv:2112.09045,

    arXiv

  5. [5]

    Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly de- tection

    Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. Adaclip: Adapting clip with hybrid learnable prompts for zero-shot anomaly de- tection. In European Conference on Computer Vision, pages 55–72. Springer, 2024. 1

    work page 2024

  6. [6]

    Center-aware residual anomaly synthesis for multiclass industrial anomaly detec- tion

    Qiyu Chen, Huiyuan Luo, Haiming Yao, Wei Luo, Zhen Qu, Chengkan Lv, and Zhengtao Zhang. Center-aware residual anomaly synthesis for multiclass industrial anomaly detec- tion. IEEE Transactions on Industrial Informatics, pages 1– 11, 2025. 1

    work page 2025

  7. [7]

    Xuhai Chen, Yue Han, and Jiangning 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. 1, 2, 3, 6, 7, 13, 16, 17, 18, 19

    arXiv 2023

  8. [8]

    Sub-image anomaly detection with deep pyramid correspondences

    Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. CoRR, abs/2005.02357, 2020. 3

    arXiv 2005

Show all 43 references
  1. [9]

    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. In Inter- national conference on pattern recognition, pages 475–489. Springer, 2021. 3

    2021

  2. [10]

    The pascal visual object classes (voc) challenge

    Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303–338, 2010. 15, 16

    2010

  3. [11]

    Fastrecon: Few-shot indus- trial anomaly detection via fast feature reconstruction

    Zheng Fang, Xiaoyang Wang, Haocheng Li, Jiejie Liu, Qi- ugui Hu, and Jimin Xiao. Fastrecon: Few-shot indus- trial anomaly detection via fast feature reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17481–17490, 2023. 1, 6, 7, 14, 18, 19

    2023

  4. [12]

    Metauas: Universal anomaly segmentation with one-prompt meta-learning

    Bin-Bin Gao. Metauas: Universal anomaly segmentation with one-prompt meta-learning. Advances in Neural Infor- mation Processing Systems, 37:39812–39836, 2025. 3, 14

    2025

  5. [13]

    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 Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 1932–1940, 2024. 1, 2, 3, 6, 7, 14, 16, 18, 19

    1932

  6. [14]

    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 , 55:216–227, 2019. 6, 11, 19

    2019

  7. [15]

    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 European Conference on Computer Vision, pages 303–319. Springer, 2022. 1, 3, 6, 7, 14, 18, 19

    2022

  8. [16]

    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 Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19606–19616, 2023. 1, 2...

    2023

  9. [17]

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

    Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex condi- tions. In 2021 13th International congress on ultra mod- ern telecommunications and control systems and w...

    2021

  10. [18]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

    arXiv

  11. [19]

    Deep learning

    Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436–444, 2015. 5

    2015

  12. [20]

    Promptad: Learn- ing 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: Learn- ing prompts with only normal samples for few-shot anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 16838– 1...

    2024

  13. [21]

    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. arXiv preprint arXiv:2402.18998, 2024. 1, 2

    2024 arXiv

  14. [22]

    Medical image classification using gen- eralized zero shot learning

    Dwarikanath Mahapatra, Behzad Bozorgtabar, and Zongyuan Ge. Medical image classification using gen- eralized zero shot learning. In Proceedings of the IEEE/CVF 9 international conference on computer vision , pages 3344– 3353, 2021. 1

    2021

  15. [23]

    From softmax to sparsemax: A sparse model of attention and multi-label clas- sification

    Andre Martins and Ramon Astudillo. From softmax to sparsemax: A sparse model of attention and multi-label clas- sification. In International conference on machine learning, pages 1614–1623. PMLR, 2016. 4

    2016

  16. [24]

    The multimodal brain tumor image segmentation benchmark (brats)

    Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (brats). IEEE transactions on medical imaging , 34(...

    1993

  17. [25]

    Vt-adl: A vision trans- former network for image anomaly detection and localiza- tion

    Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adl: A vision trans- former network for image anomaly detection and localiza- tion. In 2021 IEEE 30th International Symposium on Indus- trial Electronics (ISIE), pages 01–06. IEEE, 2...

    2021

  18. [26]

    Lscad: A large-small model collaboration framework for un- supervised industrial anomaly detection

    Shichen Qu, Xian Tao, Xinyi Gong, Zhen Qu, Mukesh Prasad, Fei Shen, Zhengtao Zhang, and Guiguang Ding. Lscad: A large-small model collaboration framework for un- supervised industrial anomaly detection. IEEE Transactions on Instrumentation and Measurement, 2025. 1

    2025

  19. [27]

    Investigating shift equivalence of convolutional neural net- works in industrial defect segmentation

    Zhen Qu, Xian Tao, Fei Shen, Zhengtao Zhang, and Tao Li. Investigating shift equivalence of convolutional neural net- works in industrial defect segmentation. IEEE Transactions on Instrumentation and Measurement, 72:1–17, 2023. 1

    2023

  20. [28]

    Vcp-clip: A visual context prompting model for zero-shot anomaly segmenta- tion

    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 segmenta- tion. In European Conference on Computer Vision , pages 301–317. Springer, 2024. 1

    2024

  21. [29]

    Bayesian prompt flow learning for zero-shot anomaly detec- tion

    Zhen Qu, Xian Tao, Xinyi Gong, Shichen Qu, Qiyu Chen, Zhengtao Zhang, Xingang Wang, and Guiguang Ding. Bayesian prompt flow learning for zero-shot anomaly detec- tion. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 30398–30408, 2025. 1

    2025

  22. [30]

    Learning transferable visual models from natural language supervi- sion

    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 supervi- sion. In International conference on machine learning, ...

    2021

  23. [31]

    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 Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318–14328, 2022. 3

    2022

  24. [32]

    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 Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1907– 1916, 2021. 2

    1907

  25. [33]

    Maeday: Mae for few-and zero-shot anomaly-detection

    Eli Schwartz, Assaf Arbelle, Leonid Karlinsky, Sivan Harary, Florian Scheidegger, Sivan Doveh, and Raja Giryes. Maeday: Mae for few-and zero-shot anomaly-detection. Computer Vision and Image Understanding , page 103958,

  26. [34]

    A hierarchical transformation-discriminating generative model for few shot anomaly detection

    Shelly Sheynin, Sagie Benaim, and Lior Wolf. A hierarchical transformation-discriminating generative model for few shot anomaly detection. In Proceedings of the IEEE/CVF inter- national conference on computer vision , pages 8495–8504,

  27. [35]

    Learning unsupervised metaformer for anomaly de- tection

    Jhih-Ciang Wu, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng- Luh Liu. Learning unsupervised metaformer for anomaly de- tection. In Proceedings of the IEEE/CVF international con- ference on computer vision, pages 4369–4378, 2021. 3

    2021

  28. [36]

    Learning unsupervised metaformer for anomaly detection

    Jhih-Ciang Wu, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng- Luh Liu. Learning unsupervised metaformer for anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4369–4378, 2021. 1

    2021

  29. [37]

    Pushing the limits of fewshot anomaly de- tection in industry vision: Graphcore

    Guoyang Xie, Jinbao Wang, Jiaqi Liu, Feng Zheng, and Yaochu Jin. Pushing the limits of fewshot anomaly de- tection in industry vision: Graphcore. arXiv preprint arXiv:2301.12082, 2023. 3

    2023 arXiv

  30. [38]

    Resad: A simple framework for class generalizable anomaly detection

    Xincheng Yao, Zixin Chen, Chao Gao, Guangtao Zhai, and Chongyang Zhang. Resad: A simple framework for class generalizable anomaly detection. Advances in Neural Infor- mation Processing Systems, 37:125287–125311, 2025. 3, 6, 7, 14, 16

    2025

  31. [39]

    Draem- a discriminatively trained reconstruction embedding for sur- face anomaly detection

    Vitjan Zavrtanik, Matej Kristan, and Danijel Skoˇcaj. Draem- a discriminatively trained reconstruction embedding for sur- face anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision , pages 8330– 8339, 2021. 5, 11, 12

    2021

  32. [40]

    Scene parsing through ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 633–641,

  33. [41]

    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 The Twelfth Interna- tional Conference on Learning Representations, 2024. 1, 13

    2024

  34. [42]

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

    Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sam- ple prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17826– 17836, 2024. 3

    2024

  35. [43]

    raw_img_path

    Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre- training for anomaly detection and segmentation. In Eu- ropean Conference on Computer Vision , pages 392–408. Springer, 2022. 6, 11, 14, 15, 16, 19 10 Appendix fo...

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.