Pith. sign in

REVIEW 3 major objections 6 minor 50 references

Defect-aware Hybrid Prompt Optimization via Progressive Tuning for Zero-Shot Multi-type Anomaly Detection and Segmentation

T0 review · 3 major / 6 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read A small set of shared learnable prompt tokens plus fixed defect-name text lets a single frozen vision-language model detect and segment seen and novel defects without per-domain training.

desk verdict DAPO's binary detection results and hybrid prompt design are real, but its zero-shot multi-type segmentation claim is overbroad and the abstract's numbers don't match the tables. read the letter →

arxiv 2512.09446 v3 pith:4DSA6IU4 submitted 2025-12-10 cs.CV

Nadeem Nazer , Hongkuan Zhou , Lavdim Halilaj , Ylli Sadikaj , Steffen Staab This is my paper · ORCID
classification cs.CV
keywords anomalydetectionsegmentationzero-shotlearningprompttuningvision-languagemodelsdistributionshiftindustrialinspectionmulti-typedefect
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 tries to show that zero-shot multi-type anomaly detection and segmentation can be handled by a single frozen vision-language model, if the model is steered with prompts that combine a few learnable context tokens with the literal name of each defect type. The proposed method, DAPO, learns these shared tokens on a source industrial dataset and then reuses them for any new defect by simply writing the defect name into the prompt. The authors report that this beats handcrafted-prompt and prompt-learning baselines on five public industrial benchmarks and an internal semiconductor dataset, improving image-level AUROC and average precision by a few points under distribution shift and improving localization of novel defect types by about 5 to 6.5 points. The practical upshot, if correct, is that manufacturers could deploy one pretrained model across many product lines and defect types without retraining or writing new prompts for each defect.

What carries the argument

The central object is a hybrid defect-aware prompt: a sequence of learnable context tokens shared across all defect types, followed by a fixed defect-name token, the word 'anomaly', and the word 'object'. Because the learnable tokens are shared, a prompt for an unseen defect is built at test time by changing only the fixed defect name. The second mechanism is progressive tuning: learnable tokens are inserted into every layer of the frozen text and vision transformers, with each layer's output connected to the next by a small decay factor, so the representation is gradually re-aligned toward defect semantics. The text embeddings are then compared with multi-stage patch embeddings through line

What would settle it

Take a dataset whose anomalies are mostly absence-type defects, keep the backbone and loss fixed, and measure pixel-level F1 for the missing class. If any text-only prompt, without object-level memory or a segmentation prior, could localize absence defects at levels comparable to visually present defects, the paper's central claim would be wrong; the reported 1.8% F1 for the missing class already indicates what a controlled comparison should look like.

Watch

Extended reading notes

Core claim

On the paper's own terms, DAPO establishes that a frozen vision-language model can be adapted to industrial anomaly detection and segmentation through a small set of shared learnable prompt tokens, with one fixed textual anchor per defect type. Instead of hand-crafting a prompt for every defect-object pair, the method learns a general abnormal context from a source dataset, then constructs a prompt for a new defect by concatenating that learned context with the defect's literal name. With progressive tuning that inserts learnable tokens into both the text and image encoders, image patches and defect-type texts are brought into a common representation. The paper reports average gains of 3.7%

Load-bearing premise

The method assumes every target defect manifests as a local visual change in image patches that can be matched to a textual description; defects defined by absence, such as a missing component, produce no such patch signal, and the paper's own experiments show the method fails on them.

Editorial extensions

If this is right

  • A single frozen vision-language backbone, adapted only by shared learnable tokens, can segment defect types unseen during training, such as flattening, rust, and stains, on a new product line by inserting the defect's name into the prompt template.
  • Image-level anomaly detection under distribution shift improves over prompt-learning and handcrafted-prompt baselines on public industrial datasets and on an internal semiconductor dataset.
  • The method can localize multiple defect types within one image, assigning each pixel to the defect type with the highest similarity score.
  • The multi-type claim is bounded: for defects that leave no local visual trace, such as missing components, the method yields very low F1 (1.8% on one real-world set), so the gains are concentrated in visually present, localizable defect types.

Reading between the lines

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

  • I would expect that adding an object-presence or memory-based prior, trained or prompted separately for absence-type defects, would close most of the residual gap, because the failure is structural: a missing object leaves a visually normal patch rather than a text-alignable anomaly signal.
  • The exploratory multi-label results suggest that the learned tokens separate 'anomalous' from 'normal' but not different defect types from one another; adding a fine-grained discrimination loss or per-defect prototypes could turn the segmenter into a reliable defect-type classifier.
  • Because performance peaks after one or two training epochs and degrades later, the progressive tokens behave like a cheap adaptation layer rather than dataset-specific learning; this hints at an online variant that adapts to a new factory line with only a handful of images.
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 / 6 minor

Summary. The paper proposes DAPO, a CLIP-based prompt optimization method for zero-shot multi-type anomaly detection/segmentation under distribution shift. DAPO combines fixed textual defect-type anchors with shared learnable tokens, optimizes these hybrid prompts on a source dataset (MVTec-AD), and then deploys them on target datasets (VisA, MPDD, MAD, Real-IAD, and an internal semiconductor dataset). The claimed contributions are a parameter-sharing design that removes per-defect prompt engineering, progressive prefix tuning in both encoders, and improved generalization for both binary and multi-type anomaly tasks. The paper reports state-of-the-art or competitive results versus MultiADS and other CLIP-based baselines, with particularly strong F1/AP gains on some datasets.

Significance. If the claims were fully supported, DAPO would be a valuable practical contribution: a single frozen CLIP backbone with shared learnable tokens and flexible text anchors could segment both seen and unseen defect types in new industrial domains without per-domain training or per-defect prompt engineering. The paper includes useful ablations (prompt length, loss weight, prefix depth, token initialization, prototype aggregation) and per-defect analyses that give insight into where the method works and where it fails. However, the central zero-shot multi-type claim is currently overbroad, and the headline improvement numbers are not consistent with the reported tables. The method appears to work well for visually localizable defects, but the paper presents it as a general solution for multi-type anomaly segmentation without that qualifier.

major comments (3)
  1. [Abstract and Table 1] The abstract claims a 6.5% average improvement in localizing novel anomaly types under zero-shot settings. This is not supported by Table 1. Averaging the four transfer settings in Table 1, DAPO is worse than MultiADS in AUROC on three of four datasets (VisA 88.9 vs 93.6; MAD-Sim 91.9 vs 92.1; MAD-Real 86.1 vs 89.2). The average F1 improvement is approximately 2.6 points and the average AP improvement approximately 2.2 points, not 6.5. Similarly, the abstract's 3.7% average image-level AUROC/AP improvement does not straightforwardly follow from Table 3's numbers (average AUROC gain over MultiADS on the three public transfer sets is about 3.3 points, AP about 3.4 points; including the internal dataset changes the aggregate differently). The authors should either reproduce the exact computation with a specific baseline set and metric, or correct the abstract and claims.
  2. [§3, §6.1, §6.2] The zero-shot multi-type segmentation claim is overbroad. The task definition in §3 includes all defect types in the target datasets, and the abstract/title do not qualify the claim. But the method aligns image patches to text prompts by construction; for 'missing' component defects there is no local visual anomaly to align. The paper's own Table 2 shows near-zero F1 for MAD-Real 'Missing' (1.8 vs MultiADS 3.7) and §6.1 attributes the VisA/MAD AUROC drops to 'missing' component anomalies. This is a structural limitation, not a tuning artifact, and it affects the core claim of general zero-shot multi-type segmentation. The paper should be re-scoped to visually localizable defect types, or the method must incorporate object-level priors; as written, the unqualified abstract and contribution statements overstate the findings.
  3. [§5.2 and Table 1] The implementation details state: 'We finetune DAPO using the test set of the MVTec-AD and tested it on other target datasets.' Training on the test split is an unusual protocol and needs clarification. If the model is trained on MVTec-AD's test portion (including annotated anomalies), the comparison with baselines whose results are 'directly reported from the original paper' may not be apples-to-apples unless the same source split and supervision were used. Please specify exactly which MVTec-AD split (train/test, and whether both normal and anomalous images) was used, and confirm that MultiADS was evaluated under the same source-data protocol.
minor comments (6)
  1. [Eq. (8)] The local loss term mixes notation: Dice(UP(\hat S[0]), Y[0]) and Dice(1 - UP(\hat S[0]), 1 - Y[0]) use the same normal-channel prediction; the multi-class dice over K+1 channels is not defined. Please clarify the exact multi-class loss formulation.
  2. [Figure 2 caption and §4] The acronym DAPO is defined as 'Defect-Aware Prompt Optimization' in the text but the Figure 2 caption calls it 'Defect-Aware Prompt Learning (DAPO).' Use one consistent name.
  3. [Eq. (10)] The parameter alpha is described as a 'small decaying parameter' but its value or schedule is never given. Please report it in the implementation details.
  4. [§5.1] The text lists 'five public datasets: MVTec-AD, VisA, MPDD, MAD, Real-IAD, and an internal dataset' — that is six datasets (five public plus one internal).
  5. [Table 3] The table reports AUPRO for several baselines but DAPO's AUPRO is lower on two of three transfer settings (e.g., VisA 84 vs MultiADS 89.7). The text says pixel-level results are 'competitive' but does not discuss this gap; a sentence acknowledging the trade-off would be useful.
  6. [§6.1] Typo: 'MuliADS' should be 'MultiADS.'

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the transfer protocol (train on MVTec-AD, evaluate on other datasets) and the fixed defect-name anchors make the reported gains non-tautological; the only overlap-based element is the use of the authors' own MultiADS as both the main baseline and the source of handcrafted-prompt framing, which is not load-bearing for DAPO's own numbers.

full rationale

The paper's central claim is that DAPO, trained on MVTec-AD with hybrid prompts that combine shared learnable tokens with fixed textual defect anchors, transfers zero-shot to other datasets and to unseen defect types. The target metrics (AUROC, AP, F1 on VisA, MPDD, MAD, Real-IAD, and the internal dataset) are never used to set prompt parameters; the learnable tokens are optimized only on the training dataset with L_global and L_local (Eq. 7-9). Thus the evaluation is not forced by construction. The defect-name placeholder <D> in Eq. (1) is a fixed, human-readable anchor, not learned from the target classes, so seeing new defect types is a genuine generalization test, not a self-definitional loop. The only overlapping-author citation is MultiADS [28], which is used as the principal baseline and as a characterization of hand-crafted prompt engineering; MultiADS values are taken from an external published paper (Table 1 states 'MultiADS results are directly reported from the original paper'). This is a real external evidence source, and DAPO's relative improvements are not derived from MultiADS' definitions or parameters. The paper does contain a substantial correctness-risk issue—the abstract claims a 3.7%/6.5% improvement while Table 1 shows DAPO is worse in AUROC on 3 of 4 datasets and Section 6.2 admits a 1.8% F1 on MAD-Real 'Missing' defects—but that is an overclaiming/scope problem, not circularity. There is also a minor self-citational framing: the paper credits MultiADS for defect-type semantics while arguing MultiADS requires manual prompts, and the authors of DAPO overlap with MultiADS' authors. However, that framing is not used to validate DAPO's mechanism; it is an external comparison, so it does not raise the circularity score above 1.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The method relies on several domain assumptions about CLIP's transferability and the visibility of defects. Multiple hyperparameters are hand-chosen or tuned on the source domain, and the best-epoch selection adds an implicit free parameter; none of these are fitted to the target test labels, so the transfer evaluation is not circular, but the hyperparameters could be overfit to the source-domain validation.

free parameters (7)
  • λ (loss weight) = 4.0
    Chosen by hand in §5.2; controls the balance between global and local losses. Ablation in Fig. 5(a) shows sensitivity.
  • learnable prompt length l = 5
    Number of learnable context tokens; set to 5 with E=10 prompts per state; ablation in Fig. 5(b) shows longer prompts reduce performance.
  • number of prompts E = 10
    Number of prompt variants per normal/abnormal state; averaged in Eq. (5)–(6); chosen empirically.
  • learnable tokens per layer (VPT) = 4
    Prefix tokens inserted per transformer layer; set in §5.2; no ablation shown for this count.
  • prefix depth N_d and α in Eq. (10) = full depth; α≪1
    Depth chosen at 'full tested depth' after non-monotonic ablation (Fig. 8); α is a small decaying parameter not precisely specified.
  • intermediate stages M = 4 (layers 6,12,18,24)
    Choice of ViT layers for patch embeddings; selected in appendix §10.1.
  • training hyperparameters = LR 0.001, batch 8, 518×518, 5 epochs, Adam
    Standard choices; also 'performance peaks around epoch 1 or 2' (appendix), so the best-epoch selection is a latent free parameter.
assumptions (4)
  • domain assumption CLIP's pre-trained vision and text encoders provide a shared semantic space where defect-type words can be aligned with visual defects.
    The entire method relies on this, stated in §4.3 and §4.4: the pre-trained feature space is 'pre-trained to capture image-text semantics' and must be adapted.
  • domain assumption Shared learnable tokens combined with fixed defect-type words yield prompts that transfer to unseen defect types.
    Central to zero-shot multi-type AS; the paper assumes the same learned tokens generalize across defect categories (§4.1). This is plausible but not proven beyond the presented experiments.
  • domain assumption Patch embeddings at ViT layers 6,12,18,24 capture local anomaly evidence.
    Used for local loss and similarity maps (§4.2.1, appendix §10.1); no proof that these layers are optimal.
  • domain assumption Progressive VPT (Eq. 10) with small α adapts encoders without destroying pre-trained features.
    The progressive connection mechanism is taken from prior work [32] and assumed to improve alignment; ablation shows it helps but does not establish the mechanism.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Defect-aware Hybrid Prompt Optimization via Progressive Tuning for Zero-Shot Multi-type Anomaly Detection and Segmentation." pith.science (2026). https://pith.science/paper/4DSA6IU4

@misc{pith2026251209446,
  author       = {Pith},
  title        = {Pith review of: Defect-aware Hybrid Prompt Optimization via Progressive Tuning for Zero-Shot Multi-type Anomaly Detection and Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4DSA6IU4}},
  note         = {Machine review of arXiv:2512.09446}
}
read the original abstract

Recent vision-language models (VLMs) like CLIP have shown impressive anomaly detection performance under significant distribution shift by utilizing high-level semantic information through text prompts. However, these models often overlook fine-grained defect cues, e.g., hole, cut, or scratch, that are essential for understanding the anomaly's nature. Moreover, the modality gap between images and text can lead to subtle visual evidence being poorly captured in textual descriptions. To address the gap, we enhance the representation of "abnormal" with structured semantics, bridging coarse anomaly signals and fine-grained defect categories. We propose a hybrid prompting mechanism that combines human-readable descriptions of defect types with learnable token embeddings. Building on these ideas, we introduce DAPO, a Defect-aware Prompt Optimization framework for zero-shot multi-type and binary anomaly detection and segmentation under distribution shift. DAPO aligns anomaly-relevant visual features with their corresponding textual semantics by learning hybrid defect-aware prompts that combine fixed textual anchors with trainable token embeddings. We conducted experiments on public benchmarks (MPDD, VisA, MVTec-AD, MAD, and Real-IAD) and an internal dataset. The results suggest that compared to the baseline models, DAPO achieves a 3.6% average improvement in AUROC and average precision metrics at the image level under distribution shift, and a 5.2% average improvement in AUROC and F1 when localizing novel anomaly types under zero-shot settings.

Figures

Figures reproduced from arXiv: 2512.09446 by the authors.

Figure 1
Figure 1. Existing methods (top) produce embeddings with in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed Defect-Aware Prompt Learning (DAPO) architecture. To enable zero-shot multi-type anomaly de [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Anomaly segmentation results from DAPO on VisA, [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (15 more)
Figure 4
Figure 4. Figure 4: Binary anomaly detection performance comparison. (a) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Hyperparameter analysis. a) weight balance [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Performance comparison of token initialization strate [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 8
Figure 8. Figure 8: Effect on injecting learnable tokens at different depth of [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 7
Figure 7. Figure 7: Ablation on the AUROC and AP with different [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 10
Figure 10. Figure 10: TPR and FPR across different threshold for anomaly [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: The visualization highlights the capsule object from the VisA dataset, while the model is trained on the MVTec-AD dataset. The first row represents the input, the second row shows the ground truth segmentation mask. The last row presents the segmentation results from …
Figure 12
Figure 12. Figure 12: The visualization highlights the Fryum object from the VisA dataset, with the model trained on MVTec-AD. As shown, DAPO successfully localizes the burnt defect, however the segmentation slightly overextends beyond the actual defect region, indicating minor over-segmen…
Figure 13
Figure 13. Figure 13: The visualization highlights the PCB4 object from the VisA dataset, with the model trained on MVTec-AD. As shown, DAPO successfully localizes both the burnt and dirt defects within the same image. However, it fails to detect a small scratch defect in the top-right cor…
Figure 14
Figure 14. Figure 14: The visualization highlights the Bottle object from the MVTec-AD dataset, with the model trained on VisA. Columns 1–2 show localization of the contamination defect, 3–4 capture the small break, while 5–6 successfully segment large break [PITH_FULL_IMAGE:figures/full_…
Figure 15
Figure 15. Figure 15: The visualization highlights the Tubes object from the MPDD dataset. The model, trained on MVTec-AD, accurately localizes the flattening defect with minimal noise, demonstrating its precision in segmenting flattening defects, as evident from [PITH_FULL_IMAGE:figures/…
Figure 16
Figure 16. Figure 16: The visualization highlights the Metal Plate object from the MPDD dataset. Columns 1–3 show accurate localization of the scratch defect, while columns 4–6 demonstrate precise segmentation of the rust defect [PITH_FULL_IMAGE:figures/full_fig_p021_16.png]
Figure 17
Figure 17. Figure 17: The visualization shows the Phone Battery object from the Real-IAD dataset. Our model accurately localizes both scratch and damage defects [PITH_FULL_IMAGE:figures/full_fig_p021_17.png]
Figure 18
Figure 18. Figure 18: The visualization highlights the SIM Card Set object from the Real-IAD dataset. DAPO successfully segments the scratch defect with clear localization. 11 [PITH_FULL_IMAGE:figures/full_fig_p021_18.png]
Figure 19
Figure 19. Figure 19: The visualization depicts the Firewood object from the Real-IAD dataset. Our model effectively localizes both scratch and damage defects, despite the irregular surface texture. 12 [PITH_FULL_IMAGE:figures/full_fig_p022_19.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 11 linked inside Pith

  1. [1]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592–9600, 2019. 6, 1

    2019

  2. [2]

    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. InEuropean Conference on Computer Vision, pages 55–72. Springer, 2024. 3, 1, 4

    2024

  3. [3]

    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. 2, 1

    arXiv 2023

  4. [4]

    Sub-image anomaly detection with deep pyramid correspondences

    N Cohen and Y Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arxiv 2020.arXiv preprint arXiv:2005.02357, 2005. 2

    arXiv 2020

  5. [5]

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

    Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: a patch distribution modeling framework for anomaly detection and localization. InInter- national conference on pattern recognition, pages 475–489. Springer, 2021. 2

    2021

  6. [6]

    Simclip: Refining image-text alignment with simple prompts for zero-/few-shot anomaly detection

    Chenghao Deng, Haote Xu, Xiaolu Chen, Haodi Xu, Xiao- tong Tu, Xinghao Ding, and Yue Huang. Simclip: Refining image-text alignment with simple prompts for zero-/few-shot anomaly detection. InProceedings of the 32nd ACM Interna- tional Conference on Multimedia, pages 1761–1770, 2024. 3

    2024

  7. [7]

    Anomaly detection via reverse distillation from one-class embedding

    Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9737–9746, 2022. 2

    2022

  8. [8]

    Generative adversarial networks.Commu- nications of the ACM, 63(11):139–144, 2020

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks.Commu- nications of the ACM, 63(11):139–144, 2020. 2

    2020

Show all 50 references
  1. [9]

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

    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 local- ization. InProceedings of the 32nd ACM International Con- ference on Multimedia, pages 2041–2049, 2024. 3

  2. [10]

    Openclip.If you use this software, please cite it as below, 7,

    Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, et al. Openclip.If you use this software, please cite it as below, 7,

  3. [11]

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

    2023

  4. [12]

    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. In2021 13th International congress on ultra mod- ern telecommunications and control systems and wo...

    2021

  5. [13]

    How can we know what language models know?Trans- actions of the Association for Computational Linguistics, 8: 423–438, 2020

    Zhengbao Jiang, Frank F Xu, Jun Araki, and Graham Neu- big. How can we know what language models know?Trans- actions of the Association for Computational Linguistics, 8: 423–438, 2020. 2

    2020

  6. [14]

    Maple: Multi-modal prompt learning

    Muhammad Uzair Khattak, Hanoona Rasheed, Muham- mad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19113–19122, 2023. 6

    2023

  7. [15]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

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

    arXiv

  8. [16]

    Auto-encoding vari- ational bayes, 2013

    Diederik P Kingma, Max Welling, et al. Auto-encoding vari- ational bayes, 2013. 2

    2013

  9. [17]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InProceedings of the IEEE/CVF international confer- ence on computer vision, pages 4015–4026, 2023. 3

    2023

  10. [18]

    Generalization and network design strate- gies.Connectionism in perspective, 19(143-155):18, 1989

    Yann LeCun et al. Generalization and network design strate- gies.Connectionism in perspective, 19(143-155):18, 1989. 2

    1989

  11. [19]

    The power of scale for parameter-efficient prompt tuning.arXiv preprint arXiv:2104.08691, 2021

    Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning.arXiv preprint arXiv:2104.08691, 2021. 2

    2021 arXiv

  12. [20]

    Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation.Neurocomputing, 618: 129122, 2025

    Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. Clipsam: Clip and sam collaboration for zero-shot anomaly segmentation.Neurocomputing, 618: 129122, 2025. 3

    2025

  13. [21]

    Dice loss for data-imbalanced nlp tasks

    Xiaoya Li, Xiaofei Sun, Yuxian Meng, Junjun Liang, Fei Wu, and Jiwei Li. Dice loss for data-imbalanced nlp tasks. arXiv preprint arXiv:1911.02855, 2019. 5

    1911 arXiv

  14. [22]

    Prefix-tuning: Optimiz- ing continuous prompts for generation.arXiv preprint arXiv:2101.00190, 2021

    Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimiz- ing continuous prompts for generation.arXiv preprint arXiv:2101.00190, 2021. 6

    2021 arXiv

  15. [23]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll´ar. Focal loss for dense object detection. InPro- ceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017. 5 9

    2017

  16. [24]

    Unsu- pervised two-stage anomaly detection.arXiv preprint arXiv:2103.11671, 2021

    Yunfei Liu, Chaoqun Zhuang, and Feng Lu. Unsu- pervised two-stage anomaly detection.arXiv preprint arXiv:2103.11671, 2021. 2

    2021 arXiv

  17. [25]

    Visualizing data using t-sne.Journal of machine learning research, 9 (Nov):2579–2605, 2008

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Journal of machine learning research, 9 (Nov):2579–2605, 2008. 3

    2008

  18. [26]

    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. InInternational conference on machine learning, p...

    2021

  19. [27]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318–14328, 2022. 2

    2022

  20. [28]

    Multiads: Defect-aware supervision for multi-type anomaly detection and segmentation in zero-shot learning.arXiv preprint arXiv:2504.06740, 2025

    Ylli Sadikaj, Hongkuan Zhou, Lavdim Halilaj, Stefan Schmid, Steffen Staab, and Claudia Plant. Multiads: Defect-aware supervision for multi-type anomaly detection and segmentation in zero-shot learning.arXiv preprint arXiv:2504.06740, 2025. 2, 3, 4, 7, 8, 1

    2025 arXiv

  21. [29]

    Two effects, one trigger: on the modality gap, object bias, and information imbalance in contrastive vision-language representation learning.arXiv preprint arXiv:2404.07983, 2024

    Simon Schrodi, David T Hoffmann, Max Argus, V olker Fis- cher, and Thomas Brox. Two effects, one trigger: on the modality gap, object bias, and information imbalance in contrastive vision-language representation learning.arXiv preprint arXiv:2404.07983, 2024. 2

    2024 arXiv

  22. [30]

    Attention guided anomaly localization in images

    Shashanka Venkataramanan, Kuan-Chuan Peng, Ra- jat Vikram Singh, and Abhijit Mahalanobis. Attention guided anomaly localization in images. InEuropean Conference on Computer Vision, pages 485–503. Springer,

  23. [31]

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

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

  24. [32]

    Progressive vi- sual prompt learning with contrastive feature re-formation

    Chen Xu, Yuhan Zhu, Haocheng Shen, Boheng Chen, Yix- uan Liao, Xiaoxin Chen, and Limin Wang. Progressive vi- sual prompt learning with contrastive feature re-formation. International Journal of Computer Vision, 133(2):511–526,

  25. [33]

    Dfr: Deep feature re- construction for unsupervised anomaly segmentation.arXiv preprint arXiv:2012.07122, 2020

    Jie Yang, Yong Shi, and Zhiquan Qi. Dfr: Deep feature re- construction for unsupervised anomaly segmentation.arXiv preprint arXiv:2012.07122, 2020. 2

    2012 arXiv

  26. [34]

    Long-horizon language-conditioned imita- tion learning for robotic manipulation.IEEE/ASME Trans- actions on Mechatronics, pages 1–12, 2025

    Xiangtong Yao, Tobias Blei, Yuan Meng, Yu Zhang, Hongkuan Zhou, Zhenshan Bing, Kai Huang, Fuchun Sun, and Alois Knoll. Long-horizon language-conditioned imita- tion learning for robotic manipulation.IEEE/ASME Trans- actions on Mechatronics, pages 1–12, 2025. 2

    2025

  27. [35]

    Bridging language and action: A survey of language-conditioned robot manip- ulation.CoRR, abs/2312.10807, 2023

    Hongkuan Zhou, Xiangtong Yao, Oier Mees, Yuan Meng, Ted Xiao, Yonatan Bisk, Jean Oh, Edward Johns, Mohit Shridhar, Dhruv Shah, Jesse Thomason, Kai Huang, Joyce Chai, Zhenshan Bing, and Alois Knoll. Bridging language and action: A survey of language-conditioned robot manip- ula...

    2023 arXiv

  28. [36]

    Language-conditioned imitation learning with base skill pri- ors under unstructured data.IEEE Robotics and Automation Letters, 9(11):9805–9812, 2024

    Hongkuan Zhou, Zhenshan Bing, Xiangtong Yao, Xiao- jie Su, Chenguang Yang, Kai Huang, and Alois Knoll. Language-conditioned imitation learning with base skill pri- ors under unstructured data.IEEE Robotics and Automation Letters, 9(11):9805–9812, 2024. 2

    2024

  29. [37]

    Predicting the road ahead: A knowledge graph based foundation model for scene under- standing in autonomous driving

    Hongkuan Zhou, Stefan Schimid, Yicong Li, Lavdim Halilaj, Xiangtong Yao, and Wei Cao. Predicting the road ahead: A knowledge graph based foundation model for scene under- standing in autonomous driving. InThe Semantic Web, pages 116–132, Cham, 2025. Springer Nature Switzerland. 2

    2025

  30. [38]

    Conditional prompt learning for vision-language mod- els

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language mod- els. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 16816–16825,

  31. [39]

    Learning to prompt for vision-language models.In- ternational Journal of Computer Vision, 130(9):2337–2348,

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models.In- ternational Journal of Computer Vision, 130(9):2337–2348,

  32. [40]

    Pad: A dataset and benchmark for pose-agnostic anomaly detection.Advances in Neural Information Processing Systems, 36:44558–44571,

    Qiang Zhou, Weize Li, Lihan Jiang, Guoliang Wang, Guyue Zhou, Shanghang Zhang, and Hao Zhao. Pad: A dataset and benchmark for pose-agnostic anomaly detection.Advances in Neural Information Processing Systems, 36:44558–44571,

  33. [41]

    Anomalyclip: Object-agnostic prompt learn- ing for zero-shot anomaly detection.arXiv preprint arXiv:2310.18961, 2023

    Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. Anomalyclip: Object-agnostic prompt learn- ing for zero-shot anomaly detection.arXiv preprint arXiv:2310.18961, 2023. 2, 4, 8, 1

    2023

  34. [42]

    Xingcheng Zhou, Mingyu Liu, Ekim Yurtsever, Bare Luka Zagar, Walter Zimmer, Hu Cao, and Alois C. Knoll. Vision language models in autonomous driving: A survey and out- look.IEEE Transactions on Intelligent Vehicles, pages 1–20,

  35. [43]

    Fine-grained abnormality prompt learning for zero- shot anomaly detection.arXiv preprint arXiv:2410.10289,

    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,

  36. [44]

    Unsupervised anomaly detection with an enhanced teacher for student- teacher feature pyramid matching

    Mohammad Zolfaghari and Hedieh Sajedi. Unsupervised anomaly detection with an enhanced teacher for student- teacher feature pyramid matching. In2022 27th Interna- tional Computer Conference, Computer Society of Iran (CS- ICC), pages 1–4. IEEE, 2022. 2

    2022

  37. [45]

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

    Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre- training for anomaly detection and segmentation. InEu- ropean Conference on Computer Vision, pages 392–408. Springer, 2022. 6, 1 10 Defect-aware Hybrid Prompt Opt...

    2022

  38. [46]

    Table 5 summarizes key statistics of these datasets, includ- ing the number of distinct product classes and the distri- bution of normal and anomalous samples

    Datasets In this section, we describe the public benchmark indus- trial datasets, MvTec-AD [1], VisA [45], MPDD [12], MAD [40], Real-IAD [31], along with an internal dataset. Table 5 summarizes key statistics of these datasets, includ- ing the number of distinct product classe...

  39. [47]

    The results of the baselines are taken directly from the respective papers

    Baselines To evaluate the performance of our DAPO, we compare DAPO with several baseline models. The results of the baselines are taken directly from the respective papers. De- tails of the respective baselines are given as follows: •CLIP-AC[26]: They use an ensemble of text p...

  40. [48]

    implementation details

    Experiments Here, we will discuss in detail on DAPO through the experi- ments, and the ablation we carried out. We will also display the visualization and fine-grained results of our approach. 10.1. Implementation Details As mentioned in the “implementation details” subsection...

  41. [49]

    contam- ination

    Prototype Aggregation To determine the robust representation of the final anomaly prototype,z Dk, We perform ablation on the aggregation method used to combine information from multiple defect prompts. Baseline: By default, we define the anomaly pro- totype prompt by calculati...

  42. [50]

    We present six examples of object from VisA, MPDD, MVTec-AD, and Real-IAD datasets

    Visualization Here, we present the visualization of our anomaly segmen- tation results. We present six examples of object from VisA, MPDD, MVTec-AD, and Real-IAD datasets. Figures 11, 12, and 13 shows the capsule, fryum and pcb from VisA. Figures 15 and 16 shows the tubes and ...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.