Pith. sign in

REVIEW 6 major objections 7 minor 1 cited by

IQE-CLIP: Instance-aware Query Embedding for Zero-/Few-shot Anomaly Detection in Medical Domain

T0 review · 6 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Instance-aware query embeddings that fuse text prompts with the test image's own visual features are better indicators of medical abnormality than text-only embeddings, and IQE-CLIP reports state-of-the-art zero- and few-shot results…

desk verdict A real attempt at medical zero-/few-shot AD with a novel query-embedding architecture, but the reported gains may be inflated by conditioning the reference embeddings on the test image itself; worth reviewing seriously but not accepting until that is controlled. read the letter →

arxiv 2506.10730 v3 pith:IJRQ2ZD2 submitted 2025-06-12 cs.CV

Hong Huang , Weixiang Sun , Zhijian Wu , Jingwen Niu , Donghuan Lu , Xian Wu , Yefeng Zheng This is my paper
classification cs.CV
keywords zero-shotanomalydetectionfew-shotmedicalCLIPqueryembeddingprompttuningcross-attentioninstance-aware
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

Most CLIP-based zero-/few-shot anomaly detectors classify an image by comparing its visual features with handcrafted or tuned text embeddings for "normal" and "abnormal." This paper argues that such text-only comparators are too coarse for medical images, because the joint embedding space does not separate normal from anomalous instances of the same category. To fix this, IQE-CLIP builds a per-image query embedding: the query is initialized from the image's own [CLS] feature, carrying category and position context, then refined by cross-attention over both the tuned text embedding and multi-stage visual features. The anomaly score and anomaly map are computed from the cosine similarity between this query and normal/abnormal variants, fused with the text-based score. On six medical datasets spanning five modalities, the paper reports the best average AUC in zero-shot and few-shot settings compared with previous CLIP-based methods, and ablations show that each prompt-tuning and query-module component contributes.

What carries the argument

The load-bearing object is the Instance-aware Query Module (IQM), a cross-attention query builder that turns a positional-class query into normal/abnormal query embeddings $F^Q \in \mathbb{R}^{2\times C}$. The positional-class query is initialized as $\mathrm{MLP}(x_{\mathrm{cls}})$ plus a position embedding, where $x_{\mathrm{cls}}$ is the [CLS] token of the input image from CLIP's frozen image encoder. Each IQM block applies self-attention to the query, then text cross-attention with the tuned text embedding $F^T$, then image cross-attention with multistage visual features $F^P_i$ projected by a query adapter, followed by a feed-forward network after each attention. The prompt-tuning half supplies the text embedding: class-based prompting tokens $\{x_i\}_{i=1}^r = \mathrm{MLP}(x_{\mathrm{cls}})$ are added to learnable tokens inside the general template $P_c = [a][photo][of][a][state][class][x_1+v_1]\cdots[x_r+v_r]$, and additional learnable prompt tokens are inserted into the first $D$ layers of the text encoder. The final score combines per-layer query anomaly maps and text anomaly maps (Eqs. 10--12), so the mechanism's job is to inject category, position, and instance-level visual context into the comparator before any cosine similarity is measured.

What would settle it

Replace the CLS-derived initialization in both the class-based prompting tokens (Eq. 3) and the positional-class query with a fixed, category-level prototype shared across all test images, keeping the IQM weights and training procedure unchanged; if zero-shot AUC on the six datasets does not drop substantially, the reported instance-aware gains come from cross-modal interaction rather than from the test image's own features, whereas a large drop would confirm that injecting the instance's class and position context is essential.

Watch

Extended reading notes

Core claim

The paper's central claim is that abnormality is better indicated by an instance-aware query embedding than by a text embedding alone. Given an image, IQE-CLIP first tunes the text side with three kinds of tokens: a generic class-agnostic template, class-based prompting tokens obtained by projecting the image's [CLS] feature through an MLP into token space, and learnable prompting tokens inserted into the first layers of the frozen text encoder. It then initializes a positional-class query from the same [CLS] feature plus position embeddings and refines it through repeated self-attention, text cross-attention, and image cross-attention, producing normal and abnormal query embeddings $F^Q_N$ and $F^Q_A$. The final anomaly map is a weighted sum of per-layer cosine-similarity maps between image features and the query pair and between image features and the text pair, with the weighting $\alpha$ set to 0.8 by default. With this design the paper reports state-of-the-art zero-shot results on most benchmark tasks and the best average few-shot AUC at K=4 over previous CLIP-based methods, with ablations showing that removing either prompt tuning or the query module lowers performance.

Load-bearing premise

A load-bearing premise is that conditioning the class-based prompts and the initial query on the test image's own [CLS] feature injects only category and position context, not the image's identity, so the normal-vs-abnormal cosine separation is not an artifact of self-similarity.

Editorial extensions

If this is right

  • Zero-shot deployment no longer needs per-category handcrafted prompt lists: a class-agnostic template plus image-derived and learned tokens adapts to a new modality or anatomical region without target-domain labels.
  • The same computation yields both an image-level anomaly score (the maximum of the fused map) and a pixel-level anomaly map, so classification and localization come from one forward pass.
  • Few-shot gains are additive: with K=2, 4, 8, or 16 labeled target images, the reported average AUC rises with shot count on most datasets, and the K=4 average exceeds the strongest prior method by about 1.8 points.
  • Ablations imply the two contributions are complementary: dropping the class-based and learnable prompt tokens or dropping the query module lowers AUC in both zero- and few-shot settings, so neither the text-side adaptation nor the instance-aware interaction is redundant.

Reading between the lines

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

  • Because the prompt template is class-agnostic and the query is built from the input image itself, nothing in the method is specific to radiology or pathology; a direct test would be to run IQE-CLIP on industrial or satellite imagery, which the authors list as future work.
  • The paper does not fully isolate how much of the gain comes from genuine instance-awareness versus self-conditioning: both the class-based prompt tokens and the query initialization are functions of the same test image's [CLS] feature, so a controlled ablation that replaces $\mathrm{MLP}(x_{\mathrm{cls}})$ with a fixed category prototype would clarify the contribution.
  • The reported $\alpha=0.8$ optimum for fusing query and text maps is an overall average; since datasets differ in modality and annotation type, a per-dataset or per-layer learned $\alpha$ could push the gains further, a testable extension the paper does not run.
  • The visual separation between normal and anomalous query embeddings suggests the query space could be used beyond cosine scoring, for instance as input to a lightweight density estimator or segmentation head; the paper only reports AUC, so that is an unexplored direction.
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

6 major / 7 minor

Summary. IQE-CLIP proposes an instance-aware query embedding framework for zero-/few-shot medical anomaly detection. It extends CLIP with three-token prompt tuning (general, class-based, and learnable tokens) and an instance-aware query module (IQM) that initializes a positional-class query from the image CLS token and refines it via self-/cross-attention with text embeddings and multistage visual features. The final anomaly map is a weighted combination of text-based and query-based cosine-similarity maps. Experiments on six BMAD datasets report state-of-the-art AUCs in zero-shot and few-shot settings, with ablations on BrainMRI.

Significance. The idea of instance-conditioned query embeddings is interesting, and the paper evaluates on a useful medical benchmark with released code. If the self-conditioning concern is resolved, the framework would provide a practical CLIP adaptation for medical zero-/few-shot anomaly detection. At present, however, the lack of a control for test-image-dependent reference embeddings, the test-set selection of alpha, and a mismatch between the training and inference weighting rules lower confidence in the claimed state-of-the-art margins.

major comments (6)
  1. [§3.3, Eq. (3) and §3.4] The class-based prompting tokens x_i = MLP(x_cls) and the positional-class query initialization both use the CLS token of the test image itself. Since the text embeddings F_T^N, F_T^A and query embeddings F_Q^N, F_Q^A used in the cosine similarities of Eq. (10) are therefore functions of the image being scored, the reported improvements over fixed text prompts may be partly a self-similarity artifact rather than evidence that query embeddings are better abnormality indicators. The paper provides no loss or architectural constraint that prevents the MLP or query adapter from relying on the exact input, and the ablations in Tables 3–5 remove entire pathways rather than isolating this effect. Please add a control experiment in which x_cls is replaced by a category-level prototype or a fixed reference image (not the test image), and report whether the gains persist; also report the distribution of cos(F_i^P, F_Q^N) for normal versus abnormal test images to show that the normal reference is not simply tracking the input.
  2. [§4.3, Figure 3] The hyperparameter alpha appears to be selected on the test set: Figure 3 presents zero-shot and few-shot AC/AS curves over alpha, and the same alpha=0.8 is then used in Tables 1–2 and in Appendix B. This leaks test information into model selection. Please clarify whether alpha was chosen on a validation split, or re-run the main comparisons with a validation-selected alpha and report the differences.
  3. [§3.5, Eq. (14) vs Eq. (12)] With alpha=0.8, the training loss in Eq. (14) weights L_text by 0.8 and L_query by 0.2, while the inference map in Eq. (12) weights M_query by 0.8 and M_text by 0.2. This asymmetry is not explained. If the equations are accurate, the training objective does not match the test-time combination. Please correct the equations or provide a rationale for the asymmetry and show that the reported results are robust to it.
  4. [§3.4, Eqs. (7)–(9)] The text describes a single positional-class query, but the final query embedding is F_Q in R^{2×C} with normal and abnormal slots F_Q^N and F_Q^A used in Eq. (10). Please specify how the two slots are initialized and how the attention operations in Eqs. (7)–(9) produce two output vectors rather than one; this is necessary for reproducing the method.
  5. [§3.5, Eq. (13) and Table 6] The loss L_i includes focal and Dice terms on pixel-level annotation G, but three of the six BMAD datasets (HIS, ChestXray, OCT17) provide only image-level labels. In the leave-one-out zero-shot protocol, the training set always contains at least one such dataset. Please specify how L_i is computed for datasets without G (for example, by dropping the pixel terms or using only BCE).
  6. [Table 2 vs Table 7] For MVFA with K=4, the RESC segmentation AUC is 98.66 in Table 2 but 90.62 in Table 7, and 90.62 is exactly the average AUC reported in Table 2. One of these numbers is a transcription error. Please correct it and audit other entries in the tables and figures for similar inconsistencies.
minor comments (7)
  1. [§4.3, Tables 3–5] The ablations are run only on BrainMRI; the general claim that each component helps across medical modalities would be better supported by ablations on at least one additional dataset with pixel annotations, such as LiverCT or RESC.
  2. [§3.3, Eq. (1)] The role of the token [class] in the 'class-agnostic' template is unclear; please state whether it is a placeholder that is replaced by a category name or a literal word, since this affects the interpretation of the class-based prompting tokens.
  3. [§3.4, Eq. (5)] The notation A_k(ReLU(F_i^P, W_k)) is not mathematically valid, as ReLU takes a single argument; please write the projection explicitly, for example A_k(ReLU(F_i^P W_k)).
  4. [§3.4, Eqs. (10)–(11)] The symbol eF_i^P is used in the cosine terms but is not defined; presumably it denotes the L2-normalized F_i^P. Please define it in the text.
  5. [§4.3, Figure 3] The alpha curves are reported without error bars or standard deviations, so the claim that alpha=0.8 always yields the best results lacks uncertainty information.
  6. [References] References [53] and [54] duplicate the same MedCLIP paper; please consolidate them.
  7. [Throughout] There are typos in the Introduction ('apporaches') and in §4.3 ('hyperparemter'); please proofread the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method's claims are empirical and benchmarked against external BMAD data; the self-conditioning of prompts and queries is a generalization risk, not a circular reduction.

full rationale

IQE-CLIP's derivation chain is not circular under the rubric's definitions. The class-based prompting tokens in Eqs. (2)-(3) are computed as xi = MLP(x_cls) from the test image's CLS token, and the query embeddings in Eqs. (7)-(9) are formed by cross-attending the same image's features and the text embeddings. This is genuine self-conditioning and a legitimate concern about inflated similarity or poor generalization, but it is not a case where a prediction is equal to its input by construction: the normal/abnormal decision in Eqs. (10)-(11) still hinges on the fixed contrasting state tokens 'normal' versus 'abnormal', and the adapters and MLPs are trained with external anomaly labels through Eqs. (13)-(14). No fitted test-set label is renamed as a prediction, no result is justified solely by a self-citation, no uniqueness theorem is imported from the authors' prior work, and no known empirical pattern is merely relabeled as a new derivation. The SOTA claims are anchored to the external BMAD benchmark and compared against prior methods in Tables 1-2, and the ablations in Tables 3-5 provide independent evidence for the components. I therefore find no significant circularity.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The central claim rests on a small set of hand-chosen hyperparameters (notably alpha), a domain-transfer assumption for CLIP, and the validity of the BMAD benchmark as a zero-shot proxy. The test-image-conditioned query is an invented internal representation, but it is not a new physical entity and has no evidence outside the paper's own evaluations.

free parameters (6)
  • fusion weight alpha = 0.8
    Balances text and query anomaly maps; selected by sweeping on test performance in Figure 3 rather than on a separate validation split.
  • prompt layers D = 11
    Number of text encoder layers with learnable prompt tokens, chosen by hand (Sec. 3.3, Appendix B).
  • learnable token length M = 4
    Number of learnable prompt tokens attached in the first D layers (Sec. 3.3, Appendix B).
  • IQM blocks N = 4
    Number of cross-attention blocks in the instance-aware query module (Appendix B).
  • number of context vectors r = not reported
    Length of the learnable context token sequence in Eq. (1); the paper does not state the value used.
  • training hyperparameters = lr 0.001, 50 epochs, batch 32
    Chosen by hand and reported in Appendix B; typical settings but not tuned on a separate validation set.
assumptions (4)
  • domain assumption CLIP's frozen image and text encoders provide a joint embedding space that is meaningful for medical images after prompt tuning.
    The entire method relies on CLIP's pretrained representations transferring to medical modalities; invoked in Sec. 3.3 and 3.4.
  • domain assumption The BMAD leave-one-out protocol is a valid measure of zero-shot anomaly detection.
    Zero-shot results are obtained by training on all other BMAD datasets and testing on the held-out modality, as described in Sec. 4.2; this is a domain-generalization benchmark, not a no-training zero-shot setting.
  • domain assumption The pixel-level and image-level annotations in the training datasets are correct and complete.
    The training loss in Sec. 3.5 uses these labels; if labels are noisy, the learned query embeddings inherit the noise.
  • domain assumption Cosine similarity after L2 normalization is an appropriate anomaly scoring function.
    Eqs. (10)-(12) define anomaly maps as softmax over cosine similarities; this is a design choice, not a derived result.
invented entities (1)
  • Instance-aware query embedding FQ
    purpose: A learned per-image reference vector that combines text and multistage visual features to score anomalies via cosine similarity.
    It is an internal representation evaluated only through the paper's own experiments; there is no external falsifiable handle outside the proposed pipeline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IQE-CLIP: Instance-aware Query Embedding for Zero-/Few-shot Anomaly Detection in Medical Domain." pith.science (2026). https://pith.science/paper/IJRQ2ZD2

@misc{pith2026250610730,
  author       = {Pith},
  title        = {Pith review of: IQE-CLIP: Instance-aware Query Embedding for Zero-/Few-shot Anomaly Detection in Medical Domain},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IJRQ2ZD2}},
  note         = {Machine review of arXiv:2506.10730}
}
read the original abstract

Recently, the rapid advancements of vision-language models, such as CLIP, leads to significant progress in zero-/few-shot anomaly detection (ZFSAD) tasks. However, most existing CLIP-based ZFSAD methods commonly assume prior knowledge of categories and rely on carefully crafted prompts tailored to specific scenarios. While such meticulously designed text prompts effectively capture semantic information in the textual space, they fall short of distinguishing normal and anomalous instances within the joint embedding space. Moreover, these ZFSAD methods are predominantly explored in industrial scenarios, with few efforts conducted to medical tasks. To this end, we propose an innovative framework for ZFSAD tasks in medical domain, denoted as IQE-CLIP. We reveal that query embeddings, which incorporate both textual and instance-aware visual information, are better indicators for abnormalities. Specifically, we first introduce class-based prompting tokens and learnable prompting tokens for better adaptation of CLIP to the medical domain. Then, we design an instance-aware query module (IQM) to extract region-level contextual information from both text prompts and visual features, enabling the generation of query embeddings that are more sensitive to anomalies. Extensive experiments conducted on six medical datasets demonstrate that IQE-CLIP achieves state-of-the-art performance on both zero-shot and few-shot tasks. We release our code and data at https://github.com/hongh0/IQE-CLIP/.

Figures

Figures reproduced from arXiv: 2506.10730 by the authors.

Figure 1
Figure 1. Visualization of text and query embeddings [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Framework of IQE-CLIP, which primarily consists of two steps: prompt tuning and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Impact of fusiong weight α. Components of IQM. In [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison with few-shot anomaly detection methods on datasets of BrainMRI, LiverCT, RESC, HIS, ChestXray, and OCT17, with the shot number K = {2, 4, 8, 16}. The average AUCs (in %) for anomaly classification (AC) and anomaly segmentation (AS) across different shot num…
Figure 6
Figure 6. Figure 6: Quantitative comparison in zero-shot set￾ting with SOTA methods on medical datasets. Left: Radar chart of pixel-level AUC on three datasets with pixel-level annotations. Right: Radar chart of image-level AUC on six datasets with image-level anomaly annotations. HIS Che…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Beyond Static Anchors: Bounded Prototype Conditioning for Language-Free Medical Anomaly Detection

    cs.CV 2026-08 conditional novelty 6.0 of 10

    An input-conditioned, bounded prototype-pair mechanism makes CLIP-based medical anomaly detection adaptive per image and sets new state-of-the-art AUROC across six medical benchmarks.

Reference graph

Works this paper leans on

66 extracted references · 52 canonical work pages · cited by 1 Pith paper

  1. [1]

    The RSNA-ASNR-MICCAI BraTS 2021 benchmark on brain tumor segmentation and radiogenomic classification.arXiv preprint arXiv:2107.02314, 2021

    Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, et al. The RSNA-ASNR-MICCAI BraTS 2021 benchmark on brain tumor segmentation and radiogenomic classification.arXiv preprint arXiv:2107.02314, 2021

    arXiv 2021

  2. [2]

    Advancing the cancer genome atlas glioma MRI collections with expert segmentation labels and radiomic features

    Spyridon Bakas, Hamed Akbari, Aristeidis Sotiras, Michel Bilello, Martin Rozycki, Justin S Kirby, John B Freymann, Keyvan Farahani, and Christos Davatzikos. Advancing the cancer genome atlas glioma MRI collections with expert segmentation labels and radiomic features. Scientific Data, 4(1):1–13, 2017

    work page 2017

  3. [3]

    BMAD: Benchmarks for medical anomaly detection

    Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxiang Zhang, and Xingyu Li. BMAD: Benchmarks for medical anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4042–4053, 2024

    work page 2024

  4. [4]

    Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer.JAMA, 318(22):2199–2210, 2017

    Babak Ehteshami Bejnordi, Mitko Veta, Van Diest, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer.JAMA, 318(22):2199–2210, 2017

    work page 2017

  5. [5]

    The MVTec anomaly detection dataset: a comprehensive real-world dataset for unsupervised anomaly detection.International Journal of Computer Vision, 129(4):1038–1059, 2021

    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(4):1038–1059, 2021

    work page 2021

  6. [6]

    The liver tumor segmentation benchmark (LiTS).Medical Image Analysis, 84:102680, 2023

    Patrick Bilic, Patrick Christ, Hongwei Bran Li, et al. The liver tumor segmentation benchmark (LiTS).Medical Image Analysis, 84:102680, 2023

    2023

  7. [7]

    Deep autoencoders for anomaly detection in textured images using CW-SSIM

    Andrea Bionda, Luca Frittoli, and Giacomo Boracchi. Deep autoencoders for anomaly detection in textured images using CW-SSIM. InInternational Conference on Image Analysis and Processing, pages 669–680, 2022

    work page 2022

  8. [8]

    Dual-distribution discrepancy with self-supervised refinement for anomaly detection in medical images.Medical Image Analysis, 86:102794, 2023

    Yu Cai, Hao Chen, Xin Yang, Yu Zhou, and Kwang-Ting Cheng. Dual-distribution discrepancy with self-supervised refinement for anomaly detection in medical images.Medical Image Analysis, 86:102794, 2023

    work page 2023

Show all 66 references
  1. [9]

    Informative knowledge distillation for image anomaly segmentation.Knowledge-Based Systems, 248:108846, 2022

    Yunkang Cao, Qian Wan, Weiming Shen, and Liang Gao. Informative knowledge distillation for image anomaly segmentation.Knowledge-Based Systems, 248:108846, 2022

    2022

  2. [10]

    Segment any anomaly without training via hybrid prompt regularization.arXiv preprint arXiv:2305.10724, 2023

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

    2023 arXiv

  3. [11]

    BiaS: Incorporating biased knowledge to boost unsupervised image anomaly localization.IEEE Transactions on Systems, Man, and Cybernetics: Systems, 54(4):2342–2353, 2024

    Yunkang Cao, Xiaohao Xu, Chen Sun, Liang Gao, and Weiming Shen. BiaS: Incorporating biased knowledge to boost unsupervised image anomaly localization.IEEE Transactions on Systems, Man, and Cybernetics: Systems, 54(4):2342–2353, 2024

    2024

  4. [12]

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

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

    2024 arXiv

  5. [13]

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

    Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. AdaCLIP: Adapting CLIP with hybrid learnable prompts for zero-shot anomaly detection. InEuropean Conference on Computer Vision, pages 55–72. Springer, 2024

    2024

  6. [14]

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

    Varun Chandola, Arindam Banerjee, and Vipin Kumar. Anomaly detection: A survey.ACM Computing Surveys, 41(3):1–58, 2009

    2009

  7. [15]

    Xuhai Chen, Yue Han, and Jiangning Zhang. APRIL-GAN: A zero-/few-shot anomaly classifi- cation and segmentation method for CVPR 2023 V AND Workshop Challenge Tracks 1&2: 1st place on zero-shot AD and 4th place on few-shot AD.arXiv preprint arXiv:2305.17382, 2023. 10

    2023 arXiv

  8. [16]

    CLIP-AD: A language-guided staged dual-path model for zero-shot anomaly detection

    Xuhai Chen, Jiangning Zhang, Guanzhong Tian, Haoyang He, Wuhao Zhang, Yabiao Wang, Chengjie Wang, and Yong Liu. CLIP-AD: A language-guided staged dual-path model for zero-shot anomaly detection. InInternational Joint Conference on Artificial Intelligence, pages 17–33. Springer, 2024

    2024

  9. [17]

    Anomaly detection via reverse distillation from one-class embed- ding

    Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embed- ding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9737–9746, 2022

    2022

  10. [18]

    ANOVL: Adapting vision-language models for unified zero-shot anomaly localization.arXiv preprint arXiv:2308.15939, 2023

    Hanqiu Deng, Zhaoxiang Zhang, Jinan Bao, and Xingyu Li. ANOVL: Adapting vision-language models for unified zero-shot anomaly localization.arXiv preprint arXiv:2308.15939, 2023

    2023 arXiv

  11. [19]

    Catching both gray and black swans: Open- set supervised anomaly detection

    Choubo Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open- set supervised anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7388–7398, 2022

    2022

  12. [20]

    Unsupervised anomaly segmentation for brain lesions using dual semantic-manifold reconstruction

    Zhiyuan Ding, Qi Dong, Haote Xu, Chenxin Li, Xinghao Ding, and Yue Huang. Unsupervised anomaly segmentation for brain lesions using dual semantic-manifold reconstruction. In International Conference on Neural Information Processing, pages 133–144. Springer, 2022

    2022

  13. [21]

    Deep learning for medical anomaly detection–a survey.ACM Computing Surveys, 54(7):1–37, 2021

    Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clinton Fookes. Deep learning for medical anomaly detection–a survey.ACM Computing Surveys, 54(7):1–37, 2021

    2021

  14. [22]

    Memorizing normality to detect anomaly: Memory- augmented deep autoencoder 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 autoencoder for unsupervised anomaly detection. InProceedings of the IEEE/CVF International Confe...

    2019

  15. [23]

    CFlow-AD: Real-time unsupervised anomaly detection with localization via conditional normalizing flows

    Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. CFlow-AD: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 98–107, 2022

    2022

  16. [24]

    DiAD: A diffusion-based framework for multi-class anomaly detection

    Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. DiAD: A diffusion-based framework for multi-class anomaly detection. InProceedings of the AAAI Conference on Artificial Intelligence, 2024

    2024

  17. [25]

    Automated segmentation of macular edema in OCT using deep neural networks.Medical Image Analysis, 55:216–227, 2019

    Junjie Hu, Yuanyuan Chen, and Zhang Yi. Automated segmentation of macular edema in OCT using deep neural networks.Medical Image Analysis, 55:216–227, 2019

    2019

  18. [26]

    Registration based few-shot anomaly detection

    Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratlin, and Yanfeng Wang. Registration based few-shot anomaly detection. InEuropean Conference on Computer Vision, pages 303–319, 2022

    2022

  19. [27]

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

    Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xinchao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11375–11385, 2024

    2024

  20. [28]

    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 Conference on Computer Vision and Pattern Recognition, pages 19606–19616, 2023

    2023

  21. [29]

    A masked reverse knowledge distillation method incorporating global and local information for image anomaly detection.Knowledge-Based Systems, page 110982, 2023

    Yuxin Jiang, Yunkang Cao, and Weiming Shen. A masked reverse knowledge distillation method incorporating global and local information for image anomaly detection.Knowledge-Based Systems, page 110982, 2023

    2023

  22. [30]

    Identifying medical diagnoses and treatable diseases by image-based deep learning.Cell, 172(5):1122–1131, 2018

    Daniel S Kermany, Michael Goldbaum, Wenjia Cai, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning.Cell, 172(5):1122–1131, 2018

    2018

  23. [31]

    MaPLe: Multi-modal prompt learning

    Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fa- had Shahbaz Khan. MaPLe: Multi-modal prompt learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113–19122, 2023. 11

    2023

  24. [32]

    Segment anything

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, et al. Segment anything. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015–4026, 2023

    2023

  25. [33]

    MICCAI multi-atlas labeling beyond the cranial vault–workshop and challenge

    Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. MICCAI multi-atlas labeling beyond the cranial vault–workshop and challenge. InPoceedings MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, volume 5, page 12, 2015

    2015

  26. [34]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9664–9674, 2021

    2021

  27. [35]

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

    Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. PromptAD: Learning prompts with only normal samples for few-shot anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16838–16848, 2024

    2024

  28. [36]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2980–2988, 2017

    2017

  29. [37]

    Real3D-AD: A dataset of point cloud anomaly detection.Advances in Neural Information Processing Systems, 36, 2024

    Jiaqi Liu, Guoyang Xie, Ruitao Chen, Xinpeng Li, Jinbao Wang, Yong Liu, Chengjie Wang, and Feng Zheng. Real3D-AD: A dataset of point cloud anomaly detection.Advances in Neural Information Processing Systems, 36, 2024

    2024

  30. [38]

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

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, et al. Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. InEuropean Conference on Computer Vision, pages 38–55. Springer, 2025

    2025

  31. [39]

    The multimodal brain tumor image segmentation benchmark (BraTS).IEEE Transactions on Medical Imaging, 34(10):1993–2024, 2014

    Bjoern H Menze, Andras Jakab, Stefan Bauer, et al. The multimodal brain tumor image segmentation benchmark (BraTS).IEEE Transactions on Medical Imaging, 34(10):1993–2024, 2014

    1993

  32. [40]

    V-Net: Fully convolutional neural networks for volumetric medical image segmentation

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-Net: Fully convolutional neural networks for volumetric medical image segmentation. InInternational Conference on 3D Vision, pages 565–571. IEEE, 2016

    2016

  33. [41]

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

    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 segmentation. In European Conference on Computer Vision, pages 301–317. Springer, 2024

    2024

  34. [42]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learning transferable visual models from natural language supervision. InInternational Conference on Machine Learning, pages 8748–8763. PMLR, 2021

    2021

  35. [43]

    Towards total recall in industrial anomaly detection

    Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14318–14328, 2022

    2022

  36. [44]

    CLIP for all things zero-shot sketch-based image retrieval, fine-grained or not

    Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. CLIP for all things zero-shot sketch-based image retrieval, fine-grained or not. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 276...

    2023

  37. [45]

    Multiresolution knowledge distillation for anomaly detection

    Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad Hossein Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14902–14912, 2021

    2021

  38. [46]

    DualCoOp: Fast adaptation to multi-label recognition with limited annotations.Advances in Neural Information Processing Systems, 35:30569–30582, 2022

    Ximeng Sun, Ping Hu, and Kate Saenko. DualCoOp: Fast adaptation to multi-label recognition with limited annotations.Advances in Neural Information Processing Systems, 35:30569–30582, 2022. 12

    2022

  39. [47]

    Deep learning for unsupervised anomaly localization in industrial images: A survey.IEEE Transactions on Instrumentation and Measurement, 71:1–21, 2022

    Xian Tao, Xinyi Gong, Xin Zhang, Shaohua Yan, and Chandranath Adak. Deep learning for unsupervised anomaly localization in industrial images: A survey.IEEE Transactions on Instrumentation and Measurement, 71:1–21, 2022

    2022

  40. [48]

    Visualizing data using t-sne.Journal of Machine Learning Research, 9(86):2579–2605, 2008

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Journal of Machine Learning Research, 9(86):2579–2605, 2008

    2008

  41. [49]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural ...

    2017

  42. [50]

    Industrial Image Anomaly Localization Based on Gaussian Clustering of Pretrained Feature.IEEE Transactions on Industrial Electronics, 69(6):6182–6192, 2022

    Qian Wan, Liang Gao, Xinyu Li, and Long Wen. Industrial Image Anomaly Localization Based on Gaussian Clustering of Pretrained Feature.IEEE Transactions on Industrial Electronics, 69(6):6182–6192, 2022

    2022

  43. [51]

    Real-IAD: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection

    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 detection. InProceedings of the IEEE/CVF Conference on Compute...

    2024

  44. [52]

    ChestX-ray8: Hospital-scale chest X-ray database and benchmarks on weakly- supervised classification and localization of common thorax diseases

    Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. ChestX-ray8: Hospital-scale chest X-ray database and benchmarks on weakly- supervised classification and localization of common thorax diseases. InProceedings of the IEEE Conference on Co...

    2017

  45. [54]

    Medclip: Contrastive learning from unpaired medical images and text.arXiv preprint arXiv:2210.10163, 2022

    Zifeng Wang, Zhenbang Wu, Dinesh Agarwal, and Jimeng Sun. Medclip: Contrastive learning from unpaired medical images and text.arXiv preprint arXiv:2210.10163, 2022

    2022 arXiv

  46. [55]

    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. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 4369–4378, 2021

    2021

  47. [56]

    AnoDDPM: Anomaly detection with denoising diffusion probabilistic models using simplex noise

    Julian Wyatt, Adam Leach, Sebastian M Schmon, and Chris G Willcocks. AnoDDPM: Anomaly detection with denoising diffusion probabilistic models using simplex noise. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 650–656, 2022

    2022

  48. [57]

    SQUID: Deep feature in-painting for unsupervised anomaly detection

    Tiange Xiang, Yixiao Zhang, Yongyi Lu, Alan L Yuille, Chaoyi Zhang, Weidong Cai, and Zongwei Zhou. SQUID: Deep feature in-painting for unsupervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23890–23901, 2023

    2023

  49. [58]

    Haiming Yao, Wenyong Yu, and Xue Wang. A feature memory rearrangement network for visual inspection of textured surface defects toward edge intelligent manufacturing.IEEE Transactions on Automation Science and Engineering, 20(4):2616–2635, 2022

    2022

  50. [59]

    Explicit boundary guided semi-push-pull contrastive learning for better anomaly detection.arXiv preprint arXiv:2207.01463, 1, 2022

    Xincheng Yao, Chongyang Zhang, and Ruoqi Li. Explicit boundary guided semi-push-pull contrastive learning for better anomaly detection.arXiv preprint arXiv:2207.01463, 1, 2022

    2022 arXiv

  51. [60]

    DSR–a dual subspace re-projection network for surface anomaly detection

    Vitjan Zavrtanik, Matej Kristan, and Danijel Sko ˇcaj. DSR–a dual subspace re-projection network for surface anomaly detection. InEuropean conference on computer vision, pages 539–554. Springer, 2022

    2022

  52. [61]

    MediCLIP: Adapting CLIP for few-shot medical image anomaly detection

    Ximiao Zhang, Min Xu, Dehui Qiu, Ruixin Yan, Ning Lang, and Xiuzhuang Zhou. MediCLIP: Adapting CLIP for few-shot medical image anomaly detection. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 458–468. Springer, 2024

    2024

  53. [62]

    Conditional prompt learning for vision-language models

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816–16825, 2022. 13

    2022

  54. [63]

    Learning to prompt for vision-language models.International Journal of Computer Vision, 130(9):2337–2348, 2022

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models.International Journal of Computer Vision, 130(9):2337–2348, 2022

    2022

  55. [64]

    Encoding structure-texture relation with P-Net for anomaly detection in retinal images

    Kang Zhou, Yuting Xiao, Jianlong Yang, Jun Cheng, Wen Liu, Weixin Luo, Zaiwang Gu, Jiang Liu, and Shenghua Gao. Encoding structure-texture relation with P-Net for anomaly detection in retinal images. InEuropean Conference on Computer Vision, pages 360–377. Springer, 2020

    2020

  56. [65]

    AnomalyCLIP: Object- agnostic prompt learning 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 learning for zero-shot anomaly detection.arXiv preprint arXiv:2310.18961, 2023

    2023

  57. [66]

    Towards high-resolution 3D anomaly detection via group-level feature contrastive learning

    Hongze Zhu, Guoyang Xie, Chengbin Hou, Tao Dai, Can Gao, Jinbao Wang, and Linlin Shen. Towards high-resolution 3D anomaly detection via group-level feature contrastive learning. InProceedings of the 32nd ACM International Conference on Multimedia, pages 4680–4689, 2024. 14 A S...

    2024

  58. [67]

    For IQM, the number of attention heads is set to 8, and the number of blocks N is set to 4

    The dimensions of the query embeddings are set to 768. For IQM, the number of attention heads is set to 8, and the number of blocks N is set to 4. The fusion weight α is set to 0.8 by default for predicting the anomaly map. The Adam optimizer is employed with a fixed learning ...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.