Pith. sign in

REVIEW 5 major objections 5 minor 36 references

NexViTAD: Few-shot Unsupervised Cross-Domain Defect Detection via Vision Foundation Models and Multi-Task Learning

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read NexViTAD claims that fusing frozen Hiera and DINOv2 features, training with multi-task pseudo-labels, and clustering a normal-image memory bank with Sinkhorn K-means transfers pixel-level defect detection to unlabeled target domains…

desk verdict NexViTAD reassembles known components into a plausible pipeline, but the reporting contradictions and undefined normal-image selection make the SOTA claim unsupported as written. read the letter →

arxiv 2507.07579 v1 pith:OINRJGT4 submitted 2025-07-10 cs.CV cs.AI

Tianwei Mu , Feiyu Duan , Bo Zhou , Dan Xue , Manhong Huang This is my paper
classification cs.CVcs.AI
keywords anomalydetectioncross-domaintransferfew-shotlearningvisionfoundationmodelsmulti-taskpseudo-labelingSinkhornK-meansMVTecAD
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

The paper tries to establish that a defect-detection model trained on labeled source domains can localize defects in a target domain for which it has no labels, using only a few unlabeled target images. It claims the right recipe is frozen foundation-model features (Hiera's multi-scale hierarchy interleaved with DINOv2's semantic embeddings) plus a multi-task decoder that transfers source supervision through confidence-filtered pseudo-labels. On MVTec AD category splits, it reports target-domain AUC of 97.5%, AP of 70.4%, and PRO of 95.2%, above XDNet and CAI baselines. If correct, it offers pixel-level cross-domain defect localization without target annotations at near-real-time inference cost.

What carries the argument

The central mechanism is a two-phase pipeline with three load-bearing pieces. First, the shared subspace encoder: frozen Hiera multi-scale features pass through a lightweight bottleneck adapter (down-projection to $r = d_{\text{hiera}}/4$, GELU, up-projection, residual skip) and are interleaved slice-by-slice along channels with DINOv2 features projected to Hiera's channel count, so each scale carries both local hierarchy and semantic abstraction. Second, the multi-task decoder: separate heads per source class plus target segmentation and pseudo-labeling heads; pseudo-labels are kept only where softmax confidence exceeds 0.7, and a consistency loss on augmented target images propagates supervision into the target domain. Third, inference: the decoder is discarded, normal target images form a memory bank, Sinkhorn K-means clusters their flattened features into 30 prototypes, each test patch is scored by its minimum Euclidean distance to a prototype, and the resulting map is upsampled and Gaussian-smoothed.

What would settle it

Choose a target split with known ground-truth masks, build the memory bank by random sampling from the full unlabeled target set so defects enter at their natural rate while keeping all other components identical, and measure AUC and PRO; a large drop would show the normality assumption carries the result, while a small drop would show tolerance to contaminated prototypes.

Watch

Extended reading notes

Core claim

NexViTAD's central claim is that combining frozen foundation-model features, multi-task pseudo-label training, and prototype-based inference overcomes domain shift in unsupervised defect detection. During training, Hiera's frozen multi-scale features are adapted by a lightweight bottleneck adapter and interleaved channel-by-channel with projected DINOv2 features, forming a shared subspace; a multi-task decoder with separate heads for each source-domain class and for the target domain supervises segmentation on the source while confidence-filtered pseudo-labels at a threshold of 0.7 plus perturbation consistency supervise the target. During inference, the decoder is removed, normal target images form a memory bank, Sinkhorn K-means clusters their features into 30 prototypes, and each test patch is scored by its minimum Euclidean distance to those prototypes, followed by upsampling and Gaussian smoothing. On MVTec AD splits, the paper reports 97.5% AUC, 70.4% AP, and 95.2% PRO on the 11/1 target setting, beating XDNet and CAI variants, with ablations attributing the gains mainly to memory-bank clustering and pseudo-labeling.

Load-bearing premise

The load-bearing premise is that defect-free target images can be identified without target labels, since the memory bank is built from 'M normal images selected from Dt' and anomaly scores are distances to their prototypes; if normality is guessed rather than known, contaminated prototypes bias every score.

Editorial extensions

If this is right

  • If the reported numbers hold, a factory could deploy defect detection to a new product line by supplying a few normal images from that line and labeled images from other lines, without collecting defect annotations for the new line.
  • The paper's ablations imply that the memory-bank Sinkhorn K-means inference is the main driver: replacing it with the trained decoder drops AUC by 10 to 20 points across splits, so prototype-based matching, not the decoder itself, localizes defects.
  • Pseudo-labeling and per-domain decoder heads improve every split, suggesting that even noisy self-supervision filtered at 0.7 confidence helps alignment without source-only rigid supervision.
  • Inference time grows roughly linearly with prototype count and stays below about 100 ms for 30 prototypes and batch size 10, so near-real-time screening is compatible with the method.
  • Performance degrades as the number of target classes grows: AUC falls only from 97.5% to 95.9% but AP falls from 70.4% to 52.7%, while PRO rises to 99.1%, which the paper reads as prioritizing spatial coverage over confidence.

Reading between the lines

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

  • The paper leaves implicit that its unsupervised setting depends on knowing which target images are normal: Eq. (10) selects M normal images for the memory bank, and the paper never states how normality is decided without labels; testing with defects deliberately placed in the memory bank would show how far this assumption carries the reported numbers.
  • The evaluation treats MVTec AD category splits as domains, which is a label shift rather than a sensor or illumination shift; true deployment domain shifts may be harsher, and the AP drop on the 8/4 split hints that precision is the first metric to suffer.
  • A natural extension, not tested in the paper, is to update the memory bank online with high-confidence normal patches from the target stream, which could reduce the dependence on a pre-selected normal set.
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

5 major / 5 minor

Summary. The paper proposes NexViTAD, a framework for few-shot unsupervised cross-domain defect detection that combines frozen Hiera and DINOv2 features through a hierarchical adapter and shared subspace projection, trains a multi-task decoder with source-domain supervision and target-domain pseudo-labeling, and performs inference by clustering normal target features into Sinkhorn K-means prototypes. On MVTec AD, the authors report state-of-the-art target-domain results: AUC 97.5%, AP 70.4%, and PRO 95.2%, with ablations across four source/target split configurations and varying cluster counts.

Significance. If the claims were reliable, the work would be a useful contribution: it integrates two vision foundation models with a lightweight adapter, avoids target-domain labels at inference, and provides a public code repository. The ablation study is reasonably broad, covering decoder configuration, pseudo-labeling, multi-task training, cluster number, and inference time. However, the central performance claim is undermined by an inconsistent evaluation protocol, a nonstandard PRO metric, an unspecified procedure for selecting normal target images, and an unverifiable baseline citation. These issues are load-bearing because the paper's headline contribution is the reported state-of-the-art numbers.

major comments (5)
  1. [§4.1 and §4.3] Section 4.1 states that each configuration is run five times and average results are reported, while Section 4.3 states that each configuration is repeated 20 times and the best values are reported. Since Tables 1 and 2 provide only single point estimates, the reported AUC/AP/PRO values cannot be interpreted unambiguously; if they are best-of-20 selections, the comparison against XDNet and CAI is optimistically biased. Please report a single consistent protocol with variance estimates.
  2. [Eq. (17)] Equation (17) defines PRO as the mean IoU over test images, which is not the standard MVTec per-region overlap metric used in the anomaly-detection literature. Consequently, the PRO values in Tables 1 and 2 are not directly comparable to previously published PRO scores, and the claimed PRO gain of 2.8% over XDNet is not supported. Recompute PRO under the standard definition or rename the metric and avoid comparisons with existing PRO numbers.
  3. [§3.2 and Eq. (10)] Equation (10) and Section 3.2 state that M normal images are selected from the unlabeled target domain Dt to form the memory bank, but no selection procedure is described. If normality is taken from ground-truth labels, the method is not fully unsupervised on the target domain; if normality is inferred, the memory bank can be contaminated and the anomaly scores biased. This is a load-bearing gap because the memory bank is the entire inference mechanism.
  4. [Table 2 and Ref. [31]] Table 2 cites reference [31] for CAI-N and CAI-M, but that reference is a materials-science paper on supercapacitor electrodes and contains no cross-domain anomaly-injection method. The baseline numbers are therefore unverifiable. Please provide the correct citation and, if the correct source has different numbers, update the comparison.
  5. [§4.4.3] The cluster count K=30 is selected in Section 4.4.3 by evaluating K in {5,10,20,30,40} on the same target-domain splits whose metrics are then advertised in Tables 1 and 2, and Section 4.3 additionally reports best-of-20 splits. This constitutes tuning on the test target domains and inflates the reported performance. Report results on held-out target splits or provide a nested selection procedure.
minor comments (5)
  1. [§3.1.3] Equation (8) uses lambda1 and lambda2, while the surrounding text refers to lambda_s and lambda_t; please align the notation.
  2. [§3.2] The text says features are clustered into 4K prototypes, but Eq. (13) sums over K prototypes; clarify whether K or 4K prototypes are used.
  3. [Figure S11] The Figure S11 caption refers to an '11/4 split' although the experiments use 11/1, 10/2, 9/3, and 8/4 splits; this appears to be a typo.
  4. [References [2] and [20]] Reference [20] is cited for EfficientAD but points to a paper by Thakran and Gupta different from the original EfficientAD work, which appears to be correctly cited in [2]; please correct the citation.
  5. [Abstract] The abstract contains the typo 'Valuated' instead of 'Evaluated'.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline SOTA numbers are selected by test-set tuning (K=30, best-of-20 splits) and the 'unsupervised' inference assumes ground-truth normality, making the central claim partially circular.

  1. fitted input called prediction [Section 4.3 (setup) and Section 4.4.3 (cluster ablation)]
    "The number of clusters is 30, and the batch size of good images is 10. Each configuration is repeated 20 times with random splits, and the best values are reported. ... The best trade-off is achieved as 30 clusters."

    The final reported performance (AUC 97.5, AP 70.4, PRO 95.2) is not an independent out-of-sample prediction: K=30 is selected in Section 4.4.3 by sweeping 5, 10, 20, 30, 40 clusters on the same target-domain splits whose metrics are then advertised as the method's performance, and Section 4.3 further reports the best values over 20 random splits rather than an average. Section 4.1 instead promises five-run averages, so the tables contain a selection maximum, not a stable estimate. The headline 'prediction' is therefore partly forced by the hyperparameter and split-selection procedure.

  2. self definitional [Section 3 (problem statement) and Section 3.2, Eq. (10)]
    "an unlabeled target domain Dt = {Ii}Nt i=1. Then it performs unsupervised inference on an unlabeled target domain Dt. [...] M normal images are selected from Dt to form a memory bank:"

    The method defines the target domain as unlabeled and claims fully unsupervised inference, but the memory bank in Eq. (10) requires knowing which target images are 'normal'. Normality is a label. If M normal images are chosen using ground-truth defect-free status, then target labels are secretly used in inference and the 'unsupervised' claim is violated. If normality is instead inferred from the same unlabeled features that are later scored, defective images can enter the memory bank and contaminate the normal prototypes. Either way, the claimed unsupervised localization depends on an unstated label assumption, so the result is partly defined by that assumption rather than derived from unlabeled data alone.

full rationale

The architecture-level derivation is not circular in a mathematical sense: the frozen Hiera/DINOv2 backbones, adapters, pseudo-labeling losses, and Sinkhorn K-means clustering are standard components assembled from external work, and no load-bearing self-citation chain is present (Ref. [14] and [29] are by other research groups, and the mis-cited Ref. [31] is a citation error rather than self-citation). The circularity is concentrated in the evaluation protocol that supports the paper's central SOTA claim. The reported target-domain numbers are selected by tuning the cluster count on the same test splits and by reporting best-of-20 random splits, which makes the 'prediction' a fitted maximum rather than an independent result. Additionally, the memory-bank construction in Eq. (10) presupposes knowledge of which unlabeled target images are normal, contradicting the stated unsupervised setting. The nonstandard PRO definition (mean IoU rather than the usual MVTec per-region overlap) and the unverifiable CAI baselines are further correctness concerns, but they are not themselves circular. Because the headline achievement depends on these protocol-level reductions, a partial-circularity score of 6 is appropriate.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a handful of hand-chosen hyperparameters, most importantly K=30 clusters tuned on the target-domain evaluation splits, the pseudo-label threshold, and unstated loss weights. The main domain assumptions are that frozen foundation-model features transfer and that the memory bank of 'normal' target images is uncontaminated. No new physical entities are introduced.

free parameters (6)
  • Number of clusters K = 30
    Selected as the best trade-off across target-domain AUC, AP, and PRO in Section 4.4.3 on the same splits used for headline results.
  • Pseudo-label confidence threshold theta = 0.7
    Chosen in Eq. 6 to filter pseudo-labels; no sensitivity analysis or external justification provided.
  • Loss weights lambda1 and lambda2 = not reported
    Introduced in Eq. 8 to balance source and target losses; values are never given, so the training objective is underspecified.
  • Memory bank size M (good images) = 10
    Set as 'batch size of good images' in Section 4.3; no analysis of sensitivity to this count.
  • Gaussian smoothing sigma = 2
    Used in Eq. 14 to smooth anomaly maps; chosen without justification.
  • Adapter bottleneck dimension r = dhiera / 4
    Architecture choice in Eq. 1; no ablation on bottleneck width.
assumptions (4)
  • domain assumption Frozen Hiera and DINOv2 pretrained features are transferable across industrial domains and can be aligned by channel interleaving.
    The whole encoder design in Section 3.1.1 assumes these frozen features remain informative after domain shift; no experiment isolates this.
  • domain assumption Memory bank of M 'normal' target images represents the defect-free distribution of the target domain.
    Eq. 10 and Section 3.2 require normal images, but the target domain is unlabeled; normality must be known or assumed, and contamination would distort prototypes.
  • domain assumption Pseudo-labels above threshold theta = 0.7 are accurate enough to train the target decoder.
    Eqs. 6 and 7 use pseudo-labels as supervision; if they are noisy, the consistency loss amplifies errors.
  • ad hoc to paper Sinkhorn K-means with 30 prototypes yields a valid normal manifold for all splits.
    K=30 is tuned on target performance in Section 4.4.3 rather than derived from data statistics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NexViTAD: Few-shot Unsupervised Cross-Domain Defect Detection via Vision Foundation Models and Multi-Task Learning." pith.science (2026). https://pith.science/paper/OINRJGT4

@misc{pith2026250707579,
  author       = {Pith},
  title        = {Pith review of: NexViTAD: Few-shot Unsupervised Cross-Domain Defect Detection via Vision Foundation Models and Multi-Task Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OINRJGT4}},
  note         = {Machine review of arXiv:2507.07579}
}
read the original abstract

This paper presents a novel few-shot cross-domain anomaly detection framework, Nexus Vision Transformer for Anomaly Detection (NexViTAD), based on vision foundation models, which effectively addresses domain-shift challenges in industrial anomaly detection through innovative shared subspace projection mechanisms and multi-task learning (MTL) module. The main innovations include: (1) a hierarchical adapter module that adaptively fuses complementary features from Hiera and DINO-v2 pre-trained models, constructing more robust feature representations; (2) a shared subspace projection strategy that enables effective cross-domain knowledge transfer through bottleneck dimension constraints and skip connection mechanisms; (3) a MTL Decoder architecture supports simultaneous processing of multiple source domains, significantly enhancing model generalization capabilities; (4) an anomaly score inference method based on Sinkhorn-K-means clustering, combined with Gaussian filtering and adaptive threshold processing for precise pixel level. Valuated on the MVTec AD dataset, NexViTAD delivers state-of-the-art performance with an AUC of 97.5%, AP of 70.4%, and PRO of 95.2% in the target domains, surpassing other recent models, marking a transformative advance in cross-domain defect detection.

Figures

Figures reproduced from arXiv: 2507.07579 by the authors.

Figure 1
Figure 1. Overview of the NexViTAD framework: (a) showing the Encoder, multi-task Decoder, and (b) inference [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the NexViTAD Encoder-Decoder framework, highlighting the Hiera adapter, DINOv2 feature [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Example heatmap output for a target domain images, with defective regions highlighted after Gaussian [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Qualitative defect detection results on MVTec AD target domains (11/1 split), containing original images, [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Impact of Decoder configuration during inference, showing the performance comparison of the original [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Effect of pseudo-labeling and Decoder type during training, showing the performance comparison with or [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Effect of MTL during training: (a) AUC, (b) AP, and (c) PRO across the 11/1, 10/2, 9/3, and 8/4 domain splits. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Effect of cluster number on AUC, AP, and PRO across splits: (a) 11/1, (b) 10/2, (c) 9/3 and (c) 8/4. [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Inference time (ms) vs. number of prototypes for different batch sizes. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 31 canonical work pages

  1. [14]

    Adapting vision foundation models for real-time ultrasound image segmentation

    Xiaoran Zhang, Eric Z Chen, Lin Zhao, Xiao Chen, Yikang Liu, Boris Maihe, James S Duncan, Terrence Chen, and Shanhui Sun. Adapting vision foundation models for real-time ultrasound image segmentation. arXiv preprint arXiv:2503.24368, 2025

    arXiv 2025

  2. [31]

    Stones from other hills can polish the jade: facile mgo-templated synthesis of co-doped nimn ldh hollow nanotubes for high-performance asymmetric supercapacitor

    Yuzhu Shi, Juanqin Xue, Yongqi Yu, Kechun Wang, Shiping Zhao, Yuchen Shi, and Changbin Tang. Stones from other hills can polish the jade: facile mgo-templated synthesis of co-doped nimn ldh hollow nanotubes for high-performance asymmetric supercapacitor. Electrochimica Acta, 535:146608, September 2025

    work page 2025

  3. [1]

    Learning to adapt structured output space for semantic segmentation

    Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kihyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, June 2018

    work page 2018

  4. [2]

    Efficientad: Accurate visual anomaly detection at millisecond- level latencies

    Kilian Batzner, Lars Heckler, and Rebecca König. Efficientad: Accurate visual anomaly detection at millisecond- level latencies. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 128–138, 2024

    work page 2024

  5. [3]

    Luc P. J. Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees G. M. Snoek, and Yuki M. Asano. Gener- alAD: Anomaly Detection Across Domains by Attending to Distorted Features, page 448–465. Springer Nature Switzerland, December 2024

    work page 2024

  6. [4]

    Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad — a comprehensive real-world dataset for unsupervised anomaly detection. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 9584–9592. IEEE, June 2019. 14

    work page 2019

  7. [5]

    DecoupleNet: Decoupled Network for Domain Adaptive Semantic Segmentation, page 369–387

    Xin Lai, Zhuotao Tian, Xiaogang Xu, Yingcong Chen, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. DecoupleNet: Decoupled Network for Domain Adaptive Semantic Segmentation, page 369–387. Springer Nature Switzerland, 2022

    work page 2022

  8. [6]

    AnomalyControl: Learning Cross-modal Semantic Features for Controllable Anomaly Synthesis

    Shidan He, Lei Liu, Xiujun Shu, Bo Wang, Yuanhao Feng, and Shen Zhao. Anomalycontrol: Learning cross-modal semantic features for controllable anomaly synthesis. arXiv preprint arXiv:2412.06510, 2024

    work page Pith review arXiv 2024

Show all 36 references
  1. [7]

    Advancing industrial object detection through domain adaptation: A solution for industry 5.0

    Zainab Fatima, Shehnila Zardari, and Muhammad Hassan Tanveer. Advancing industrial object detection through domain adaptation: A solution for industry 5.0. Actuators, 13(12):513, December 2024

    2024

  2. [8]

    Test-time feature caching network for cross-domain multilayer ceramic capacitors defect detection

    Chia-Yu Hsu and Yi-Wei Lu. Test-time feature caching network for cross-domain multilayer ceramic capacitors defect detection. Applied Soft Computing, 179:113380, July 2025

    2025

  3. [9]

    Cross-domain graph level anomaly detection

    Zhong Li, Sheng Liang, Jiayang Shi, and Matthijs van Leeuwen. Cross-domain graph level anomaly detection. IEEE Transactions on Knowledge and Data Engineering, 36(12):7839–7850, December 2024

    2024

  4. [10]

    YOLO-Pdd: A Novel Multi-scale PCB Defect Detection Method Using Deep Representations with Sequential Images, page 297–313

    Bowen Liu, Dongjie Chen, and Xiao Qi. YOLO-Pdd: A Novel Multi-scale PCB Defect Detection Method Using Deep Representations with Sequential Images, page 297–313. Springer Nature Singapore, 2025

    2025

  5. [11]

    Cross-domain few-shot anomaly detection for equipment in nuclear power plants

    Junjie He, Sheng Zheng, Shuang Yi, Senquan Yang, and Zhihe Huan. Cross-domain few-shot anomaly detection for equipment in nuclear power plants. Nuclear Engineering and Design, 436:113956, May 2025

    2025

  6. [12]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  7. [13]

    Hiera: A hierarchical vision transformer without the bells-and-whistles

    Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International conference on machine learning...

    2023

  8. [15]

    U-Net: Convolutional Networks for Biomedical Image Segmentation, page 234–241

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation, page 234–241. Springer International Publishing, 2015

    2015

  9. [16]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), page 770–778. IEEE, June 2016

    2016

  10. [17]

    A survey on foundation-model-based industrial defect detection

    Tianle Yang, Luyao Chang, Jiadong Yan, Juntao Li, Zhi Wang, and Ke Zhang. A survey on foundation-model-based industrial defect detection. arXiv preprint arXiv:2502.19106, 2025

    2025 arXiv

  11. [18]

    Towards total recall in industrial anomaly detection

    Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 14298–14308. IEEE, June 2022

    2022

  12. [19]

    Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows

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

    2021 arXiv

  13. [20]

    Efficientad: A deep learning approach for multi-stage ad classification using transfer learning

    Archana Thakran and Yogesh Kumar Gupta. Efficientad: A deep learning approach for multi-stage ad classification using transfer learning. In 2023 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS), page 1100–1106. IEEE, November 2023

    2023

  14. [21]

    Simplenet: A simple network for image anomaly detection and localization

    Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection and localization. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 20402–20411. IEEE, June 2023

    2023

  15. [22]

    SuperSimpleNet: Unifying Unsupervised and Supervised Learning for Fast and Reliable Surface Defect Detection, page 47–65

    Blaž Rolih, Matic Fuˇcka, and Danijel Skoˇcaj. SuperSimpleNet: Unifying Unsupervised and Supervised Learning for Fast and Reliable Surface Defect Detection, page 47–65. Springer Nature Switzerland, December 2024

    2024

  16. [23]

    Negative selection algorithm with constant detectors for anomaly detection

    Dong Li, Shulin Liu, and Hongli Zhang. Negative selection algorithm with constant detectors for anomaly detection. Applied Soft Computing, 36:618–632, November 2015

    2015

  17. [24]

    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. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, June 2021

    2021

  18. [25]

    DrÆm – a discriminatively trained reconstruction embedding for surface anomaly detection

    Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. DrÆm – a discriminatively trained reconstruction embedding for surface anomaly detection. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), page 8310–8319. IEEE, October 2021. 15

    2021

  19. [26]

    Anomaly Detection with Conditioned Denoising Diffusion Models, page 181–195

    Arian Mousakhan, Thomas Brox, and Jawad Tayyub. Anomaly Detection with Conditioned Denoising Diffusion Models, page 181–195. Springer Nature Switzerland, 2025

    2025

  20. [27]

    Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection

    Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 16699–16708. IEEE, June 2024

    2024

  21. [28]

    Springer Nature Switzerland, November 2024

    Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang.A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localization, page 37–54. Springer Nature Switzerland, November 2024

    2024

  22. [29]

    Xdnet: A few-shot meta-learning approach for cross-domain visual inspection

    Xian Yeow Lee, Lasitha Vidyaratne, Mahbubul Alam, Ahmed Farahat, Dipanjan Ghosh, Teresa Gonzalez Diaz, and Chetan Gupta. Xdnet: A few-shot meta-learning approach for cross-domain visual inspection. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop...

    2023

  23. [30]

    Detect everything with few examples.arXiv preprint arXiv:2309.12969, 2023

    Xinyu Zhang, Yuhan Liu, Yuting Wang, and Abdeslam Boularias. Detect everything with few examples.arXiv preprint arXiv:2309.12969, 2023

    2023 arXiv

  24. [32]

    Anomalydino: Boosting patch-based few-shot anomaly detection with dinov2

    Simon Damm, Mike Laszkiewicz, Johannes Lederer, and Asja Fischer. Anomalydino: Boosting patch-based few-shot anomaly detection with dinov2. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), page 1319–1329. IEEE, February 2025

    2025

  25. [33]

    Joint forecasting of source-load-price for integrated energy system based on multi-task learning and hybrid attention mechanism

    Ke Li, Yuchen Mu, Fan Yang, Haiyang Wang, Yi Yan, and Chenghui Zhang. Joint forecasting of source-load-price for integrated energy system based on multi-task learning and hybrid attention mechanism. Applied Energy, 360:122821, April 2024

    2024

  26. [34]

    Multitask learning

    Rich Caruana. Multitask learning. Machine learning, 28:41–75, 1997

    1997

  27. [35]

    Multi-task learning for thyroid nodule segmentation with thyroid region prior

    Haifan Gong, Guanqi Chen, Ranran Wang, Xiang Xie, Mingzhi Mao, Yizhou Yu, Fei Chen, and Guanbin Li. Multi-task learning for thyroid nodule segmentation with thyroid region prior. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), page 257–261. IEEE, April 2021

    2021

  28. [36]

    Lightspeed computation of optimal transportation distances.Advances in Neural Information Processing Systems, 26(2):2292–2300, 2013

    M Cuturi. Lightspeed computation of optimal transportation distances.Advances in Neural Information Processing Systems, 26(2):2292–2300, 2013. 16 6 Appendix Figure S1: Qualitative defect detection results on MVTec AD target domains (8/4 Split), containing original images, grou...

    2013

Pith tools

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