Pith. sign in

REVIEW 2 major objections 1 minor 156 references

Current multimodal anomaly detection models condition decisions only superficially on textual input.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.3

2026-06-28 15:02 UTC pith:LC42FTQU

load-bearing objection The paper's TGAD benchmark and APD dataset show text conditioning in anomaly detection is mostly superficial, with clear metric drops, but APD construction risks confounds that weaken the main claim. the 2 major comments →

arxiv 2606.01992 v1 pith:LC42FTQU submitted 2026-06-01 cs.CV cs.AIcs.LG

A Structured Benchmark for Text-Guided Anomaly Detection: When Language Stops Conditioning the Decision

Stefano Samele , Eugenio Lomurno , Teodora Jovanovic , Sanjay Shivakumar Manohar , Alberto Crivellaro , Matteo Matteucci This is my paper
classification cs.CV cs.AIcs.LG
keywords text-guided anomaly detectionvision-language modelsindustrial anomaly detectionbenchmark evaluationMVTec ADAssembled Panel Datasetmultimodal inspection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper develops the TGAD benchmark to test whether language actually guides anomaly detection decisions or whether reported performance stems from pretrained visual features alone. It evaluates models across three scenarios of rising language dependence: prompt sensitivity on MVTec AD, component-tagged instructions that must restrict assessment to a specified part, and the new Assembled Panel Dataset that requires both defect-type and component-location knowledge. In each case the models absorb prompt wording when it is redundant but lose discriminative power once text must actively limit the decision space. The results indicate that existing evaluation protocols cannot separate genuine text guidance from strong visual priors. The authors conclude that a protocol isolating language's functional role is required before these systems can support reliable language-controlled industrial inspection.

Core claim

In the prompt-sensitivity scenario the generative model's image-level AUROC falls from 97.4 to 82.6 when the object noun is removed. In the component-tagged scenario performance drops from 90.3 to 66.3 once defects outside the instructed part are treated as normal. On the APD dataset that combines both requirements, image-level discrimination collapses to 71.2, 50.5 and 31.5, with one model falling below chance. These patterns hold across generative large vision-language, training-free discriminative and embedding-adaptive discriminative models.

What carries the argument

The TGAD benchmark and its three progressive scenarios together with the APD dataset that forces models to apply both defect-type and component-location constraints from text.

Load-bearing premise

The new scenarios and APD dataset isolate the functional role of language without introducing confounding factors from dataset construction, prompt phrasing or model-specific preprocessing.

What would settle it

A model that maintains high image-level AUROC on APD while correctly restricting anomaly scores to only the instructed component and only the specified defect types would falsify the claim that language conditions decisions only superficially.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Standard benchmarks inherited from unimodal tasks overstate the text-guided capabilities of current multimodal anomaly detection systems.
  • Performance gains observed on conventional protocols largely reflect pretrained visual features rather than functional use of language.
  • Reliable control through language for industrial deployment requires evaluation protocols that measure whether text actually constrains the decision.
  • All three evaluated model paradigms exhibit the same superficial conditioning on textual input.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same superficial language use may appear in other text-conditioned vision tasks that rely on pretrained multimodal models.
  • Training objectives that explicitly penalize failure to follow spatial or type constraints could be tested as a direct response to the observed gaps.
  • The APD dataset could serve as a testbed for measuring whether future models improve on both defect-type and location instructions simultaneously.
  • Hybrid approaches that combine language with explicit spatial masks may be required until language conditioning becomes robust.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 1 minor

Summary. The paper introduces the Text-Guided Anomaly Detection (TGAD) benchmark with three progressively more demanding scenarios to test whether language functionally conditions anomaly decisions in multimodal models. On MVTec AD it measures prompt sensitivity (object-noun removal) and component restriction; on the new Assembled Panel Dataset (APD) it combines defect-type and location instructions. One representative model is evaluated per paradigm (generative LVLM, training-free discriminative, embedding-adaptive discriminative). The central empirical claim is that conditioning is only superficial, supported by reported I-AUROC drops (97.4→82.6, 90.3→66.3) and APD collapses (71.2/50.5/31.5, one below chance and below MVTec levels).

Significance. If the isolation of language's role holds after controls, the work is significant for showing that existing benchmarks inherited from unimodal settings overstate text-guided capabilities. The structured scenarios and APD provide a concrete, falsifiable protocol that could guide development of reliably controllable industrial systems. The concrete metric drops constitute a clear, reproducible signal that future models must surpass.

major comments (2)
  1. [APD evaluation] APD results paragraph (the 71.2/50.5/31.5 figures): the claim that combined defect-type + component-location instructions cause decision collapse rests on APD construction. No ablation is reported on (a) alternative prompt phrasings that preserve identical semantics, (b) whether APD normal labels were assigned at the same component granularity used at test time, or (c) preprocessing/normalization differences versus MVTec. Any of these alone could reproduce the observed drops, undermining attribution to superficial text conditioning.
  2. [Evaluation setup] Model selection and evaluation protocol: only a single model is tested per paradigm. Because the central claim is that 'in all three' paradigms the textual interface conditions decisions only superficially, the absence of additional models within each paradigm leaves open the possibility that the observed behavior is model-specific rather than paradigm-general.
minor comments (1)
  1. [Abstract and results] Abstract and results sections omit data-split details, number of runs, and statistical significance tests for the reported AUROC differences; adding these would strengthen reproducibility without altering the central claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We appreciate the referee's thorough review and constructive feedback. The comments raise valid points about potential confounds in the APD results and the scope of model evaluation. We address each major comment below, indicating planned revisions where appropriate.

read point-by-point responses

  1. Referee: [APD evaluation] APD results paragraph (the 71.2/50.5/31.5 figures): the claim that combined defect-type + component-location instructions cause decision collapse rests on APD construction. No ablation is reported on (a) alternative prompt phrasings that preserve identical semantics, (b) whether APD normal labels were assigned at the same component granularity used at test time, or (c) preprocessing/normalization differences versus MVTec. Any of these alone could reproduce the observed drops, undermining attribution to superficial text conditioning.

    Authors: We agree these controls would strengthen attribution of the observed drops specifically to superficial text conditioning. The current manuscript describes APD construction but does not include the requested ablations. In the revised version we will add: (a) results using alternative prompt phrasings that preserve identical semantics for defect type and location, (b) explicit verification that normal labels were assigned at the component granularity matching test-time instructions, and (c) a direct comparison of preprocessing and normalization pipelines between APD and MVTec AD. These additions will help isolate the contribution of the textual conditions. revision: yes

  2. Referee: [Evaluation setup] Model selection and evaluation protocol: only a single model is tested per paradigm. Because the central claim is that 'in all three' paradigms the textual interface conditions decisions only superficially, the absence of additional models within each paradigm leaves open the possibility that the observed behavior is model-specific rather than paradigm-general.

    Authors: The manuscript explicitly selects one representative model per paradigm (generative LVLM, training-free discriminative, embedding-adaptive discriminative) to demonstrate the pattern. We acknowledge that this does not exhaustively test every model within each paradigm and therefore cannot fully rule out model-specific effects. In revision we will add a dedicated limitations paragraph clarifying the representative nature of the selection, noting that the consistent superficial-conditioning pattern across the three distinct paradigms supports the broader claim while leaving open the value of wider model testing in follow-up work. revision: partial

Circularity Check

0 steps flagged

Empirical benchmark paper with no circular derivation chain

full rationale

The paper presents a new structured benchmark (TGAD) and dataset (APD) consisting of three progressive scenarios that empirically test the functional role of language in anomaly detection models. All central claims (performance drops from 97.4 to 82.6, 90.3 to 66.3, and 71.2/50.5/31.5) are direct results of new evaluations on held-out test conditions and a newly constructed dataset. No equations, parameter fitting, predictions derived from fitted inputs, or load-bearing self-citations appear in the derivation; the results are obtained by running representative models under controlled prompt and label variations. The work is self-contained against external benchmarks because the observed gaps are measured on the authors' own constructed test sets rather than reducing to prior fitted quantities or self-referential definitions.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The paper relies on standard computer-vision evaluation practices and introduces one new dataset; no free parameters or invented physical entities are described.

axioms (1)
  • standard math Standard image-level AUROC and similar metrics appropriately quantify anomaly detection performance under text guidance.
    Metrics are used to report all performance drops in the three scenarios.
invented entities (1)
  • Assembled Panel Dataset (APD) no independent evidence
    purpose: Realistic industrial setting requiring both defect-type and component-location knowledge for text-guided evaluation.
    New dataset created for the third scenario; no independent evidence outside the paper is provided.

pith-pipeline@v0.9.1-grok · 5875 in / 1305 out tokens · 25129 ms · 2026-06-28T15:02:46.149826+00:00 · methodology

0 comments
read the original abstract

Industrial anomaly detection has historically been a unimodal task. Recent multimodal vision-language models have produced systems that admit textual input alongside the image and are presented as enabling text-guided zero- and few-shot inspection. Yet these methods are evaluated with protocols inherited from unimodal benchmarks that hold the textual condition constant and therefore cannot measure whether language conditions the decision; whether reported gains reflect text guidance or strong pretrained visual features remains open. We introduce Text-Guided Anomaly Detection (TGAD), a structured benchmark that progressively increases the functional role of language across three scenarios: a controlled prompt-sensitivity setting on MVTec AD; a component-tagged extension of MVTec AD that requires the model to restrict its assessment to an instructed part; and the new Assembled Panel Dataset (APD), a realistic industrial setting that requires both defect-type and component-location knowledge. We evaluate one representative model per paradigm: generative large vision-language, training-free discriminative, and embedding-adaptive discriminative. In all three, the textual interface conditions the decision only superficially: prompt content is absorbed unless the object noun is removed (the generative model's I-AUROC drops from 97.4 to 82.6); component-level instructions do not constrain the decision once defects outside the instructed part are admitted as normal (from 90.3 to 66.3); and when both combine on APD, image-level discrimination collapses below the MVTec level, in one case below chance (71.2, 50.5, 31.5). These results suggest that standard benchmarks overstate the text-guided capabilities of current multimodal anomaly detection systems, and that a protocol of this kind is a prerequisite for models that can be reliably controlled through language for industrial deployment.

Figures

Figures reproduced from arXiv: 2606.01992 by Alberto Crivellaro, Eugenio Lomurno, Matteo Matteucci, Sanjay Shivakumar Manohar, Stefano Samele, Teodora Jovanovic.

Figure 1
Figure 1. Figure 1: Overview of the three TGAD scenarios, ordered by the functional role language must play in the decision. In each, a fixed image and a textual request are processed by a multimodal anomaly detection model and scored against the ground truth. Scenario 1 varies the prompt on MVTec AD to test whether the output moves with language; Scenario 2 instructs inspection of a single component, admitting defects in oth… view at source ↗
Figure 2
Figure 2. Figure 2: Example of the Assembled Panel Dataset. (a) overall panel layout; (b)–(e) close-ups of representative defects: swapped power-cable colors at a power port, an unplugged wire at an EV port, a missing red protective cap, and a missing screw on a PCB. AD) and where appropriate in Scenario 3 (per-error-type prompts on Assembled Panel); for Scenario 1 the unmod￾ified LogSAD pipeline is used. AA-CLIP. We trained … view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative cross-model comparison on the cable class under EV2, for a one-part request to inspect only the grey sub-cable. Columns: input image (a), ground truth (b), and the anomaly maps of AnomalyGPT (c), LogSAD (d), and AA-CLIP (e). Top row: a defect on the non-instructed outer insulation, to which every model responds; bottom row: a defect on the instructed grey wire, which all models localize correct… view at source ↗
Figure 4
Figure 4. Figure 4: Assembled Panel in the I-AUROC–AUPRO plane (best configuration per model); arrows run from Standard AD to Per-Error AD. The shaded band marks the level of a random classifier. 4.3 Scenario 3: Industrial Robustness on Assembled Panel Assembled Panel realizes the Scenario 1 and Scenario 2 requirements at once: the panel holds numerous interact￾ing components whose anomalies are subtle deviations in presence,… view at source ↗
Figure 5
Figure 5. Figure 5: Per-model I-AUROC (solid) and AUPRO (dashed) across the three scenarios; the shaded band is the gap between them. Under the component instruction of Scenario 2 the decision drops below localization for all three models (the Localization-Decision Dissociation); on Assembled Panel localization stays above the decision only for AA-CLIP. Scenario 3 uses Per-Error AD ( [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

156 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    Variational autoen- coder based anomaly detection using reconstruction probability.Special lecture on IE, 2(1):1–18, 2015

    Jinwon An and Sungzoon Cho. Variational autoen- coder based anomaly detection using reconstruction probability.Special lecture on IE, 2(1):1–18, 2015

    2015

  2. [2]

    Explainable artificial intelligence (xai): Concepts, taxonomies, opportunities and challenges toward responsible ai

    Alejandro Barredo Arrieta, Natalia Díaz-Rodríguez, Javier Del Ser, Adrien Bennetot, Siham Tabik, Al- berto Barbado, Salvador García, Sergio Gil-López, Daniel Molina, Richard Benjamins, et al. Explainable artificial intelligence (xai): Concepts, taxonomies, opportunities and challenges toward responsible ai. Information fusion, 58:82–115, 2020

    2020

  3. [3]

    Multimodal machine learning: A survey and taxonomy.IEEE transactions on pattern analysis and machine intelligence, 41(2):423–443, 2019

    Tadas Baltrušaitis, Chaitanya Ahuja, and Louis- Philippe Morency. Multimodal machine learning: A survey and taxonomy.IEEE transactions on pattern analysis and machine intelligence, 41(2):423–443, 2019

    2019

  4. [5]

    URLhttp://arxiv.org/abs/1807.02011

    work page internal anchor Pith review Pith/arXiv arXiv

  5. [6]

    Uninformed students: Student- teacher anomaly detection with discriminative latent embeddings

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student- teacher anomaly detection with discriminative latent embeddings. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 4183–4192, 2020

    2020

  6. [7]

    The mvtec anomaly detection dataset: A comprehensive real- world dataset for unsupervised anomaly detection

    Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. The mvtec anomaly detection dataset: A comprehensive real- world dataset for unsupervised anomaly detection. International journal of computer vision, 129:1038– 1059, 2021

    2021

  7. [8]

    Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization.International journal of computer vision, 130(4):947–969, 2022

    Paul Bergmann, Kilian Batzner, Michael Fauser, David Sattlegger, and Carsten Steger. Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization.International journal of computer vision, 130(4):947–969, 2022

    2022

  8. [9]

    Anomaly detection: A survey.ACM computing surveys, 41(3):1–58, 2009

    Varun Chandola, Arindam Banerjee, and Vipin Ku- mar. Anomaly detection: A survey.ACM computing surveys, 41(3):1–58, 2009

    2009

  9. [10]

    Aunifiedanomalysynthesisstrategywith gradient ascent for industrial anomaly detection and localization

    Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zheng- taoZhang. Aunifiedanomalysynthesisstrategywith gradient ascent for industrial anomaly detection and localization. InEuropean Conference on Computer Vision, pages37–54.Springer, 2024. Codeavailable athttps://github.com/cqylunlun/GLASS

    2024

  10. [11]

    Progressive boundary guided anomaly synthesis for industrial anomaly detection

    Qiyu Chen, Huiyuan Luo, Han Gao, Chengkan Lv, and Zhengtao Zhang. Progressive boundary guided anomaly synthesis for industrial anomaly detection. IEEETransactionsonCircuitsandSystemsforVideo Technology, 35(2):1193–1208, 2025. Codeavailable athttps://github.com/cqylunlun/PBAS

    2025

  11. [12]

    Memorizing normality to detect anomaly: Memory-augmented deep au- toencoder for unsupervised anomaly detection

    Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton Van Den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep au- toencoder for unsupervised anomaly detection. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 1705–1714, 2019

    2019

  12. [13]

    Anomalygpt: Detecting industrial anomalies using large vision- language models

    Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. Anomalygpt: Detecting industrial anomalies using large vision- language models. InProceedings of the AAAI con- ference on artificial intelligence, volume 38, pages 1932–1940, 2024

    1932

  13. [14]

    Evaluating anomaly de- tection algorithms: A multi-metric analysis across variable class imbalances

    Mohammad-Sahadet Hossain, Moham- mad Sakhawat Hossain, Simon Klüttermann, and Emmanuel Müller. Evaluating anomaly de- tection algorithms: A multi-metric analysis across variable class imbalances. In2024 International Joint Conference on Neural Networks (IJCNN), pages 1–7. IEEE, 2024

    2024

  14. [15]

    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. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 19606–19616, 2023

    2023

  15. [16]

    Mmad: A comprehensive bench- mark for multimodal large language models in indus- trial anomaly detection

    Xi Jiang, Jian Li, Hanqiu Deng, Yong Liu, Bin-Bin Gao, Yifeng Zhou, Jialin Li, Chengjie Wang, and Feng Zheng. Mmad: A comprehensive bench- mark for multimodal large language models in indus- trial anomaly detection. InInternational conference on learning representations, volume 2025, pages 87273–87295, 2025

    2025

  16. [17]

    Cutpaste: Self-supervised learning for anomaly detection and localization

    Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9664–9674, 2021

    2021

  17. [18]

    Aa-clip: Enhancingzero-shotanomaly 10 A Structured Benchmark for Text-Guided Anomaly Detection detection via anomaly-aware clip

    Wenxin Ma, Xu Zhang, Qingsong Yao, Fenghe Tang, Chenxu Wu, Yingtai Li, Rui Yan, Zihang Jiang, and SKevinZhou. Aa-clip: Enhancingzero-shotanomaly 10 A Structured Benchmark for Text-Guided Anomaly Detection detection via anomaly-aware clip. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4744–4754, 2025

    2025

  18. [19]

    Vt-adl: A vision transformer network for image anomaly detec- tion and localization

    Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Piciarelli, and Gian Luca Foresti. Vt-adl: A vision transformer network for image anomaly detec- tion and localization. In2021 IEEE 30th International Symposium on Industrial Electronics (ISIE), pages 01–06. IEEE, 2021

    2021

  19. [20]

    Deep learning for anomaly detection: A review.ACM computing surveys, 54(2): 1–38, 2021

    GuansongPang,ChunhuaShen,LongbingCao,and Anton Van Den Hengel. Deep learning for anomaly detection: A review.ACM computing surveys, 54(2): 1–38, 2021

    2021

  20. [21]

    Learning transferable visual models from natural lan- guage supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh,GabrielGoh, SandhiniAgarwal, GirishSas- try,AmandaAskell,PamelaMishkin,JackClark,etal. Learning transferable visual models from natural lan- guage supervision. InInternational conference on machine learning, pages 8748–8763. PMLR, 2021

    2021

  21. [22]

    To- wards total recall in industrial anomaly detection

    Karsten Roth, Latha Pemula, Joaquin Zepeda, Bern- hard Schölkopf, Thomas Brox, and Peter Gehler. To- wards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 14318– 14328, 2022

    2022

  22. [23]

    Deep one-class classification

    Lukas Ruff, Nico Görnitz, Lucas Deecke, Shoaib Ahmed Siddiqui, Robert Vandermeulen, Alexander Binder, Emmanuel Müller, and Marius Kloft. Deep one-class classification. InInternational conference on machine learning, pages 4393–4402. PMLR, 2018

    2018

  23. [24]

    Anomaly detec- tion using autoencoders with nonlinear dimensional- ityreduction

    Mayu Sakurada and Takehisa Yairi. Anomaly detec- tion using autoencoders with nonlinear dimensional- ityreduction. InProceedingsoftheMLSDA20142nd Workshop on Machine Learning for Sensory Data Analysis, pages 4–11, 2014

    2014

  24. [25]

    Patchwise sparse dictionary learning from pre-trained neural network activation maps for anomaly detection in im- ages

    Stefano Samele and Matteo Matteucci. Patchwise sparse dictionary learning from pre-trained neural network activation maps for anomaly detection in im- ages. In2022 26th International Conference on Pat- tern Recognition (ICPR), pages 1307–1313. IEEE, 2022

    2022

  25. [26]

    Scaling anomaly detection with segmenta- tion models.Digital Signal Processing, 168:105613, 2026

    Stefano Samele, Francesco Attorre, and Matteo Mat- teucci. Scaling anomaly detection with segmenta- tion models.Digital Signal Processing, 168:105613, 2026

    2026

  26. [27]

    Towards zero-shot anomaly detection and reasoning with multimodal large lan- guage models

    Jiacong Xu, Shao-Yuan Lo, Bardia Safaei, Vishal M Patel, and Isht Dwivedi. Towards zero-shot anomaly detection and reasoning with multimodal large lan- guage models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 20370–20382, 2025

    2025

  27. [28]

    Draem: A discriminatively trained reconstruction em- bedding for surface anomaly detection

    Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem: A discriminatively trained reconstruction em- bedding for surface anomaly detection. InProceed- ings of the IEEE/CVF international conference on computer vision, pages 8330–8339, 2021

    2021

  28. [29]

    Towards training-free anomaly detection with vision and language foundation models

    Jinjin Zhang, Guodong Wang, Yizhou Jin, and Di Huang. Towards training-free anomaly detection with vision and language foundation models. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15204–15213, 2025

    2025

  29. [30]

    Eiad: Explainableindustrialanomaly detection via multi-modal large language models

    Zongyun Zhang, Jiacheng Ruan, Xian Gao, Ting Liu, andYuzhuoFu. Eiad: Explainableindustrialanomaly detection via multi-modal large language models. In 2025 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6. IEEE, 2025

    2025

  30. [31]

    Is there any anomaly in the image?

    Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the- difference self-supervised pre-training for anomaly detection and segmentation. InEuropean confer- ence on computer vision, pages 392–408. Springer, 2022. 11 A Structured Benchmark for Text-Guided Anomaly Detection A Per-Part MVTec AD Extension: Full Annotation Specificati...

    2022

  31. [32]

    Focus only on examining the light blue insulator of the conductor for any flaws

  32. [33]

    Assess solely the light blue insulation within the conductor for signs of damage

  33. [34]

    cable2 (green cable)

    Pay close attention only to the insulation layer in the conductor, which is light blue, for any potential issues. cable2 (green cable)

  34. [35]

    Focus solely on the green wire to check for any issues

  35. [36]

    Verify the integrity of the green cable only before proceeding

  36. [37]

    cable3 (blue cable)

    Examine the green cable only in detail to ensure it’s in good condition. cable3 (blue cable)

  37. [38]

    Focus solely on the blue cable for any imperfections during inspection

  38. [39]

    Attention must be directed exclusively towards the blue cable when examining for flaws

  39. [40]

    cable4 (gray cable)

    During the examination process, ensure that you closely examine the blue cable only without looking at the other cables. cable4 (gray cable)

  40. [41]

    Focus exclusively on the gray wire when examining it for potential flaws

  41. [42]

    It is crucial to scrutinize solely the gray wire within this setting

  42. [43]

    cable1_cable2

    For the purposes of analysis, concentrate solely on the gray cable in this particular scene. cable1_cable2

  43. [44]

    For a thorough examination, focus solely on the light blue coating and the green wire in this image

  44. [45]

    When conducting a meticulous inspection, pay close attention to the light-blue insulation and the green cable depicted here

  45. [46]

    cable1_cable3

    To ensure the quality of these components, closely examine only the green wire and the light-blue coating shown in the image. cable1_cable3

  46. [47]

    Focus solely on the light blue cover and blue wire, looking out for any imperfections or faults

  47. [48]

    Examine closely solely the light-blue insulation layer and the blue cable for any potential issues

  48. [49]

    cable2_cable3

    Pay close attention only to the light blue covering and the blue conductor within the container, searching for signs of defect or malfunction. cable2_cable3

  49. [50]

    Pay attention solely to the green wire and blue wire in this image for any potential issues

  50. [51]

    Examine carefully only the green wire and the blue wire within the image for any sign of defects

  51. [52]

    cable2_cable4

    Look closely only at the green cable and the blue cable visible in the photo for any flaws or imperfections. cable2_cable4

  52. [53]

    Examine just the green wire and silver wire for potential flaws

  53. [54]

    Look closely only at the green conductor and gray tube to ensure their quality

  54. [55]

    cable3_cable4

    Check solely the green line and gray pipe for any imperfections. cable3_cable4

  55. [56]

    Focus solely on examining the blue and gray cables for possible flaws

  56. [57]

    Attention should be confined to inspecting the blue and gray cable lines only for any defects

  57. [58]

    cable2_cable3_cable4

    Check only the blue and gray cables for potential issues during the inspection process. cable2_cable3_cable4

  58. [59]

    Pay attention solely to the green cable, blue cable, and gray cable when examining the wires

  59. [60]

    Focus on evaluating only the condition of the green cable, blue cable, and gray cable for any faults

  60. [61]

    capsule1 (white printed “500”)

    Examine exclusively the green cable, blue cable, and gray cable for possible defects. capsule1 (white printed “500”)

  61. [62]

    Ensure to scrutinize solely the 500-word content for any imperfections

  62. [63]

    17 A Structured Benchmark for Text-Guided Anomaly Detection

    Check only for flaws in the lines of white text. 17 A Structured Benchmark for Text-Guided Anomaly Detection

  63. [64]

    capsule2 (left black part)

    Review specifically the 500-word content marked by white characters. capsule2 (left black part)

  64. [65]

    Take note to scrutinize just the black portion on the left for any signs of imperfection

  65. [66]

    It is crucial to examine solely the black section on the left when assessing for flaws

  66. [67]

    capsule3 (right reddish-orange part)

    Be vigilant and carefully check only the black area located on the left side for any defects. capsule3 (right reddish-orange part)

  67. [68]

    Examine solely the red/orange section on the right for any flaws

  68. [69]

    Assess just the proper hue of red/orange on the right for any imperfections

  69. [70]

    capsule1_capsule3

    Check exclusively the accurate right reddish-orange area for any defects. capsule1_capsule3

  70. [71]

    Examine just the white text and orange side of the pill for any irregularities or imperfections

  71. [72]

    Look closely only at the small white label and the larger reddish-orange portion of the pill for potential issues

  72. [73]

    capsule2_capsule3

    Focus exclusively on the 500 white writing and the right orange side of the capsule for defects that may be present. capsule2_capsule3

  73. [74]

    Look closely solely at the left black section and the right reddish-orange portion of the pill for any imperfections

  74. [75]

    Examine the pill by focusing just on the dark left side and the vibrant red right side for any flaws

  75. [76]

    hazelnut1 (shell)

    Investigate both ends of the pill, the left black part and the right orange part for any defects or issues. hazelnut1 (shell)

  76. [77]

    Check just the outer casing for any flaws before consumption

  77. [78]

    Examine solely the hard outer shell for imperfections when assessing quality

  78. [79]

    hazelnut2 (hilum)

    Take a close look only at the exterior shell to ensure it’s in good condition. hazelnut2 (hilum)

  79. [80]

    Inspect only the hilum area, do not consider the outer shell

  80. [81]

    Focus solely your attention on the circular scarred base of the nut, ignoring the shell surface

Showing first 80 references.