Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Self-Navigated Residual Mamba for Universal Industrial Anomaly Detection

T0 review · 4 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 SNARM claims that a second self-referential matching pass sets the best reported anomaly results on MVTec AD, MVTec 3D, and VisA across four training regimes.

desk verdict Solid architecture and ablations, but the headline SOTA claim is undercut by the paper's own tables and a VisA inconsistency. read the letter →

arxiv 2508.01591 v2 pith:Y4LZ27JX submitted 2025-08-03 cs.CV

Hanxi Li , Jingqi Wu , Lin Yuanbo Wu , Mingliang Li , Deyin Liu , Jialie Shen , Chunhua Shen This is my paper
classification cs.CV
keywords anomalydetectionindustrialinspectionMambaself-referentialmatchingresiduallearninguniversalmulti-classfew-shot
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

SNARM is a framework for industrial anomaly detection built around a two-stage matching idea: each test patch is first compared against a memory bank of normal training features to get an inter-residual, and then re-compared against patches inside the same test image that the first pass judged most normal, producing an intra-residual. The paper's central claim is that this self-referential second pass amplifies subtle defects and improves anomaly discrimination, while also allowing a much smaller memory bank. The combined residual features are processed by a Mamba-based decoder whose scanning paths are steered by the residual map, and sixteen directional and multi-scale views are averaged at inference. Extensive experiments report best-known numbers on MVTec AD, MVTec 3D, and VisA under multi-class, cross-class, single-class, and few-shot settings, all with a single model and 68.3 million parameters.

What carries the argument

The load-bearing object is the hybrid residual representation $\hat R = [R \parallel \tilde R]$, formed by concatenating the inter-residual $R$ (powered absolute difference between a test patch and its nearest neighbour in a coreset-subsampled normal feature bank, with exponent $\theta \in \{1,2\}$) and the intra-residual $\tilde R$ (the same operation against the $p\%$ of test patches with the lowest Waypoint Map scores). The Waypoint Map, computed from a $1\times1$ convolution plus channel-averaged residual magnitude, both selects the trusted internal references and routes the Self-Navigated Mamba Block's four directional scans toward informative regions. Each directional output is passed through atrous convolution branches with rates $\{3,6,12,24\}$, and the resulting sixteen anomaly maps are averaged at inference to produce the final prediction.

What would settle it

Construct or collect test images in which more than 25 percent of the patches are anomalous, so that with the default $p=75$ threshold the trusted reference set must contain abnormal patches, and check whether SNARM's intra-matching still localizes the defects as well as it does on small-anomaly images; a large drop in pixel AUROC or PRO would confirm that the waypoint-selection assumption in Eq. 7 is load-bearing.

Watch

Extended reading notes

Core claim

The central discovery the authors advance is that self-referential residuals carry information that residuals against an external normal-feature bank do not. After the first matching stage produces a coarse anomaly map (the Waypoint Map), the patches with the lowest waypoint scores are treated as trusted normal references drawn from the test image itself; rematching every patch against these references yields second-order residuals that highlight deviations from the image's own dominant appearance. Concatenating the inter- and intra-residuals into a hybrid residual, and feeding that hybrid into a Mamba module whose four directional scans are navigated by the Waypoint Map, is what the paper credits for its reported gains. In the paper's own terms, this hybrid matching unifies global normality deviation with refined self-referenced difference and sets new performance records across all three benchmarks and all four evaluation protocols.

Load-bearing premise

The load-bearing premise is that the test patches the first pass judges most normal really are normal; if a defect covers a large enough region, those selected reference patches will include the defect and the second matching pass will not amplify the anomaly signal.

Editorial extensions

If this is right

  • A single trained SNARM model can be applied to product categories never seen during training, since the cross-class experiments report strong results without retraining.
  • The memory bank can shrink from $10^5$ to $10^4$ prototypes with little change in scores, so the added self-reference stage does not sacrifice the memory efficiency of patch-based detectors.
  • A sequential Mamba decoder, when navigated by residual signals, can match or exceed the reported scores of transformer-based universal anomaly detectors on these benchmarks.
  • Detection and localization improve together: on multi-class MVTec AD the paper reports 79.0 pixel average precision, 96.6 PRO, 99.1 pixel AUROC, and 99.4 image AUROC.

Reading between the lines

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

  • The self-referential matching principle is domain-agnostic, so a natural extension would be video or medical imaging where normal content dominates each frame or scan; the paper does not test this.
  • The fixed percentile threshold $p=75$ is a design choice rather than a learned parameter, so an adaptive or iteratively refined waypoint threshold could extend the method to images that are mostly anomalous.
  • Because intra-matching constructs references from the test image itself at inference time, SNARM could serve as a building block for test-time adaptation when the appearance of normal parts drifts after deployment; this is an implication of the method, not a result reported in the paper.
  • A falsifiable prediction from the paper's own logic is that ablating the intra-matching stage should hurt most on categories with subtle or low-contrast defects, since those are precisely the cases the second pass is designed to amplify.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes SNARM, a universal industrial anomaly detection framework based on two-stage hybrid matching (inter-residual matching against a compact training feature bank, followed by intra-residual matching against low-residual patches selected from the test image), a self-navigated Mamba block whose scanning paths are guided by a predicted waypoint map, a multi-view decoder, and an ensemble of 4x4 directional and scale branches. Experiments on MVTec AD, MVTec 3D, and VisA under multi-class, cross-class, single-class, and few-shot protocols report strong pixel-level numbers and a favorable efficiency-accuracy trade-off. The authors claim state-of-the-art performance across all metrics and benchmarks.

Significance. The core idea of using in-image self-references to generate second-order residuals is interesting and, if validated, would be a meaningful extension of residual-based anomaly detection; the self-navigated scanning mechanism also offers a concrete way to reduce the token cost of Mamba-style decoders. The paper is an empirical benchmark contribution with a public code link, an ablation study, and detailed per-class tables, which are commendable. However, the headline claim of universal state-of-the-art performance is not supported by the paper's own tables, and at least one reported table conflicts with the appendix for the same configuration, so the empirical contribution needs correction before its significance can be assessed.

major comments (4)
  1. [Table 1 vs. Appendix Table 12] The same configuration SNARM(T=10^5) is reported with VisA multi-class scores 55.8/94.7/99.1/98.1 in Table 1 but 55.5/94.0/99.1/97.6 in Appendix Table 12, and the prose margins of 2.6/0.2/10.2 in the section 'Results under the Multi-Class Setting' match neither set of numbers. Because the central claim is an empirical benchmark record, the reader must be able to trust these tables; please correct the discrepancy and state which numbers are official.
  2. [Abstract, Conclusion, Table 1] The abstract and conclusion claim improvements in 'all metrics' and 'new performance records across MVTec AD, MVTec 3D, and VisA', but Table 1 itself contradicts this: on MVTec-AD multi-class I-AUROC is 99.4 for SNARM(T=10^5) versus 99.6 for Dinomaly and 99.7 for INP-Former, and on VisA it is 98.1 versus 98.7 and 98.9. The claims should be revised to describe competitive or state-of-the-art pixel-level metrics, not all metrics.
  3. [Eq. (7)] The assumption that test patches with the lowest waypoint scores are reliable normal references is load-bearing for Intra-Matching but is not validated. If anomalies cover large contiguous regions or subtle defects produce small inter-residuals, the reference set S will contain abnormal patches, and the intra-residuals will encode a comparison of abnormal patches to other abnormal patches rather than amplifying the anomaly signal. The ablation in Table 5 only shows that adding the component helps on MVTec-AD; it does not examine failure cases. Please add an analysis of large-area and subtle-anomaly categories or a principled guard against this failure mode.
  4. [Table 1, prose after Table 1] The statement that reducing the memory bank size has 'minimal impact' is not supported by Table 1. Comparing T=10^4 with T=10^5, MVTec-3D P-AP drops from 63.6 to 59.5, PRO from 97.4 to 96.5, and I-AUROC from 93.9 to 90.2; on VisA, PRO drops from 94.7 to 90.1 and I-AUROC from 98.1 to 96.8. These are substantial degradations and should be discussed accurately rather than described as minimal.
minor comments (5)
  1. [Table 2] The reported VisA cross-class margin of 16.5 in P-AP does not match Table 2: HETMM is 38.4 and SNARM(T=10^4) is 56.0, a difference of 17.6; please verify all margin calculations.
  2. [Table 3] In the VisA block of Table 3, the Dinomaly row lists only three metrics; please add the missing I-AUROC value or explain why it is omitted.
  3. [Eq. (7)] The notation 'lowest p%' is ambiguous: with p=75 it could mean the lowest 75% or the 75th-percentile threshold; please clarify the selection rule in the text.
  4. [Introduction/Related Work] The sentence citing 'works (Liu et al. 2021; Guo et al. 2024; He et al. 2024b)' groups Swin Transformer with two anomaly-detection papers; please rephrase so that the non-AD reference is not described as an IAD method.
  5. [Throughout] The table headers use inconsistent spacing such as 'P AP' and 'I AUROC'; use consistent hyphenated metric names throughout for readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No definitional circularity: SNARM's benchmark claims are empirical and its components are not fitted to their own outputs.

full rationale

The paper's central claim is empirical: SNARM reports anomaly detection performance on MVTec AD, MVTec 3D, and VisA across multi-class, cross-class, single-class, and few-shot settings. The method is a concrete architecture: pretrained DINOv2 features are matched against a coreset memory bank to form inter-residuals (Eqs. 4-5), low-residual patches are selected by Eq. 7 to form intra-residuals (Eq. 8), the residuals are concatenated (Eq. 9), processed by a self-navigated Mamba module and a multi-view decoder, and the final map is an average over 16 branches (Eq. 16). No parameter is fitted to the reported benchmark metric and then renamed as a prediction; the loss in Eq. 12 supervises the Residual Navigator and decoder against ground-truth anomaly maps, and the final anomaly map is not used to define any of the model's inputs. The self-citations to CPR and WeakREST appear only as comparison baselines in the tables and are not used to justify the framework, so they are not load-bearing. The assumption in Eq. 7 that low-waypoint-score patches are normal references is a modeling assumption and a potential limitation, but it is not circular: the waypoint map is derived from inter-residuals, while the final prediction is an ensemble of trained decoder outputs evaluated against external labels. The table inconsistencies noted by the reviewer are accuracy/reproducibility concerns, not instances of the derivation reducing to its own inputs. Therefore no significant circularity is present.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

No new physical or mathematical entities are introduced. The proposed components (Waypoint Map, Self-Navigated Mamba Block, Multi-View Decoder) are architectural constructs evaluated only as part of the full system; they carry no independent falsifiable handle outside the reported benchmarks. The free parameters are standard hyperparameters, but the paper does not describe a validation protocol, so their selection may have used the test benchmarks.

free parameters (8)
  • p (percentile for normal patch selection) = 75
    Patches below the 75th percentile of waypoint scores are used as in-image references in Eq. 7. Performance depends on this threshold, and no validation protocol is reported.
  • k (top-k neighbors for feature averaging) = 3
    Number of nearest bank features averaged to form a stable reference in Eq. 14.
  • lambda (consistent feature jittering scale) = 30
    Noise strength in Eq. 15; must be set by hand and affects residual stability.
  • theta (residual power exponent) = 1 or 2
    Power applied to absolute differences in Eqs. 5 and 8 to adjust anomaly contrast.
  • T (coreset bank size) = 10^4, 10^5
    Memory bank size after coreset sampling; main tables report both values and results vary with T.
  • K (cyclic training steps) = 100
    Number of sequential branch updates in the cyclic optimization that promotes diversity.
  • Focal loss hyperparameters = alpha_nav=0.5, gamma_nav=4; alpha_branch=0.25, gamma_branch=4
    Two separate settings for the Residual Navigator and view-specific branch losses; no sensitivity analysis is provided.
  • Atrous dilation rates = 3, 6, 12, 24
    Fixed multi-scale rates in the Multi-View Decoder to capture anomalies of different scales.
assumptions (5)
  • domain assumption DINOv2-R pretrained features generalize to industrial defect detection.
    The entire pipeline operates on features from a model pretrained on natural images; if these features do not separate defects from normal surfaces, the residual bank and self-referential matching fail. The paper cites Darcet et al. 2023 but provides no new evidence for this transfer.
  • domain assumption Coreset sampling preserves the information needed for anomaly detection.
    The compact bank B replaces the full pool B_raw via coreset selection in Eq. 3, following PatchCore. This assumes the sampled prototypes retain discriminative references.
  • domain assumption Small inter-residual norm indicates normality.
    Used to define the Waypoint Map and to select trusted patches in Eq. 7; this is a heuristic, not a proven property of DINOv2 features.
  • domain assumption In-image low-waypoint patches serve as valid normal prototypes.
    Intra-matching in Eq. 8 assumes these patches are anomaly-free; if they contain anomalies, the self-referential residuals are contaminated.
  • domain assumption Directional Mamba scans plus atrous convolutions capture the relevant anomaly context.
    The SMB and MVD architecture is asserted to improve localization; no theoretical or mechanistic proof is given beyond ablation results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Navigated Residual Mamba for Universal Industrial Anomaly Detection." pith.science (2026). https://pith.science/paper/Y4LZ27JX

@misc{pith2026250801591,
  author       = {Pith},
  title        = {Pith review of: Self-Navigated Residual Mamba for Universal Industrial Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y4LZ27JX}},
  note         = {Machine review of arXiv:2508.01591}
}
read the original abstract

In this paper, we propose Self-Navigated Residual Mamba (SNARM), a novel framework for universal industrial anomaly detection that leverages ``self-referential learning'' within test images to enhance anomaly discrimination. Unlike conventional methods that depend solely on pre-trained features from normal training data, SNARM dynamically refines anomaly detection by iteratively comparing test patches against adaptively selected in-image references. Specifically, we first compute the ``inter-residuals'' features by contrasting test image patches with the training feature bank. Patches exhibiting small-norm residuals (indicating high normality) are then utilized as self-generated reference patches to compute ``intra-residuals'', amplifying discriminative signals. These inter- and intra-residual features are concatenated and fed into a novel Mamba module with multiple heads, which are dynamically navigated by residual properties to focus on anomalous regions. Finally, AD results are obtained by aggregating the outputs of a self-navigated Mamba in an ensemble learning paradigm. Extensive experiments on MVTec AD, MVTec 3D, and VisA benchmarks demonstrate that SNARM achieves state-of-the-art (SOTA) performance, with notable improvements in all metrics, including Image-AUROC, Pixel-AURC, PRO, and AP.

Figures

Figures reproduced from arXiv: 2508.01591 by the authors.

Figure 1
Figure 1. Illustration of the proposed Self-Navigated [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed SNARM method. The query image is matched with an external feature bank to generate [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the proposed Self-Navigated Mamba Block (SMB). Guided by the Waypoint Map, the block se￾lectively scans informative regions using four directional passes, enabling efficient and spatially-aware message prop￾agation over reduced token sets. Better viewed in color. To address this, we propose the Self-Navigated Mamba Block (SMB), which dynamically selects a compact and informative subset of tokens based on… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative results of our SNARM on MVTec-AD under the Multi-class setting. Two SOTA methods (Dinomaly (Guo [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 25 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bergmann, P.; Fauser, M.; Sattlegger, D.; and Steger, C. 2019. MVTec AD--A comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9592--9600

    2019

  4. [4]

    Bergmann, P.; Jin, X.; Sattlegger, D.; and Steger, C. 2021. The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization. arXiv preprint arXiv:2112.09045

    arXiv 2021

  5. [5]

    Chen, Y.; Tian, Y.; Pang, G.; and Carneiro, G. 2022. Deep one-class classification via interpolated gaussian descriptor. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 383--392

    work page 2022

  6. [6]

    Chen, Z.; Xie, X.; Yang, L.; and Lai, J.-H. 2024. Hard-Normal Example-Aware Template Mutual Matching for Industrial Anomaly Detection. International Journal of Computer Vision, 1--23

    work page 2024

  7. [7]

    Damm, S.; Laszkiewicz, M.; Lederer, J.; and Fischer, A. 2025. Anomalydino: Boosting patch-based few-shot anomaly detection with dinov2. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 1319--1329. IEEE

    work page 2025

  8. [8]

    Darcet, T.; Oquab, M.; Mairal, J.; and Bojanowski, P. 2023. Vision transformers need registers. arXiv preprint arXiv:2309.16588

    arXiv 2023

Show all 45 references
  1. [9]

    Defard, T.; Setkov, A.; Loesch, A.; and Audigier, R. 2021. Padim: a patch distribution modeling framework for anomaly detection and localization. In Pattern Recognition. ICPR International Workshops and Challenges: Virtual Event, January 10--15, 2021, Proceedings, Part IV, 475...

    2021

  2. [10]

    Deng, H.; and Li, X. 2022. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9737--9746

    2022

  3. [11]

    Ding, C.; Pang, G.; and Shen, C. 2022. Catching both gray and black swans: Open-set supervised anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7388--7398

    2022

  4. [12]

    Gu, A.; and Dao, T. 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752

    2023 arXiv

  5. [13]

    Gu, A.; Goel, K.; and R \'e , C. 2021. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396

    2021 arXiv

  6. [14]

    Gu, Z.; Zhu, B.; Zhu, G.; Chen, Y.; Tang, M.; and Wang, J. 2023. AnomalyGPT : Detecting Industrial Anomalies using Large Vision - Language Models . ArXiv:2308.15366 [cs]

    2023 arXiv

  7. [15]

    Guo, J.; Lu, S.; Zhang, W.; and Li, H. 2024. Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection. arXiv e-prints

    2024

  8. [16]

    He, H.; Bai, Y.; Zhang, J.; He, Q.; Chen, H.; Gan, Z.; Wang, C.; Li, X.; Tian, G.; and Xie, L. 2024 a . Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv e-prints

    2024

  9. [17]

    He, H.; Zhang, J.; Chen, H.; Chen, X.; Li, Z.; Chen, X.; Wang, Y.; Wang, C.; and Xie, L. 2024 b . A diffusion-based framework for multi-class anomaly detection. In AAAI, volume 38, 8472--8480

    2024

  10. [18]

    Huang, C.; Guan, H.; Jiang, A.; Zhang, Y.; Spratling, M.; and Wang, Y.-F. 2022. Registration based few-shot anomaly detection. In Computer Vision--ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part XXIV, 303--319. Springer

    2022

  11. [19]

    Im, J.; Son, Y.; and Hong, J. H. 2025. FUN-AD: Fully Unsupervised Learning for Anomaly Detection with Noisy Training Data. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 9447--9456. IEEE

    2025

  12. [20]

    Iqbal, N.; and Martinel, N. 2025. Pyramid-based Mamba Multi-class Unsupervised Anomaly Detection. arXiv preprint arXiv:2504.03442

    2025 arXiv

  13. [21]

    Jeong, J.; Zou, Y.; Kim, T.; Zhang, D.; Ravichandran, A.; and Dabeer, O. 2023. Winclip: Zero-/few-shot anomaly classification and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 19606--19616

    2023

  14. [22]

    Jiang, X.; Liu, J.; Wang, J.; Nie, Q.; Wu, K.; Liu, Y.; Wang, C.; and Zheng, F. 2022. Softpatch: Unsupervised anomaly detection with noisy data. Advances in Neural Information Processing Systems, 35: 15433--15445

    2022

  15. [23]

    Li, C.-L.; Sohn, K.; Yoon, J.; and Pfister, T. 2021. Cutpaste: Self-supervised learning for anomaly detection and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 9664--9674

    2021

  16. [24]

    Li, H.; Hu, J.; Li, B.; Chen, H.; Zheng, Y.; and Shen, C. 2023 a . Target before shooting: Accurate anomaly detection and localization under one millisecond via cascade patch retrieval. arXiv preprint arXiv:2308.06748

    2023 arXiv

  17. [25]

    Y.; Chen, H.; Liu, D.; Wang, M.; and Wang, P

    Li, H.; Wu, J.; Wu, L. Y.; Chen, H.; Liu, D.; Wang, M.; and Wang, P. 2023 b . Industrial Anomaly Detection and Localization Using Weakly-Supervised Residual Transformers. arXiv preprint arXiv:2306.03492

    2023 arXiv

  18. [26]

    Li, X.; Zhang, Z.; Tan, X.; Chen, C.; Qu, Y.; Xie, Y.; and Ma, L. 2024. PromptAD: Learning Prompts with only Normal Samples for Few-Shot Anomaly Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 16838--16848

    2024

  19. [27]

    Lin, T.-Y.; Goyal, P.; Girshick, R.; He, K.; and Doll \'a r, P. 2017. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, 2980--2988

    2017

  20. [28]

    Liu, Y.; Tian, Y.; Zhao, Y.; Yu, H.; Xie, L.; Wang, Y.; Ye, Q.; Jiao, J.; and Liu, Y. 2024. Vmamba: Visual state space model. Advances in neural information processing systems, 37: 103031--103063

    2024

  21. [29]

    Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; and Guo, B. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, 10012--10022

    2021

  22. [30]

    Liu, Z.; Zhou, Y.; Xu, Y.; and Wang, Z. 2023. SimpleNet: A Simple Network for Image Anomaly Detection and Localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 20402--20411

    2023

  23. [31]

    Luo, W.; Cao, Y.; Yao, H.; Zhang, X.; Lou, J.; Cheng, Y.; Shen, W.; and Yu, W. 2025. Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2025

  24. [32]

    Pei, X.; Huang, T.; and Xu, C. 2024. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977

    2024 arXiv

  25. [33]

    Roth, K.; Pemula, L.; Zepeda, J.; Sch \"o lkopf, B.; Brox, T.; and Gehler, P. 2022 a . Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 14318--14328

    2022

  26. [34]

    Roth, K.; Pemula, L.; Zepeda, J.; Sch \"o lkopf, B.; Brox, T.; and Gehler, P. 2022 b . Towards Total Recall in Industrial Anomaly Detection . ArXiv:2106.08265 [cs]

    2022 arXiv

  27. [35]

    Russakovsky, O.; Deng, J.; Su, H.; Krause, J.; Satheesh, S.; Ma, S.; Huang, Z.; Karpathy, A.; Khosla, A.; Bernstein, M.; et al. 2015. Imagenet large scale visual recognition challenge. International journal of computer vision, 115: 211--252

    2015

  28. [36]

    V.; and Mahalanobis, A

    Venkataramanan, S.; Peng, K.-C.; Singh, R. V.; and Mahalanobis, A. 2020. Attention guided anomaly localization in images. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XVII, 485--503. Springer

    2020

  29. [37]

    Wang, F.; Wang, J.; Ren, S.; Wei, G.; Mei, J.; Shao, W.; Zhou, Y.; Yuille, A.; and Xie, C. 2024. Mamba-r: Vision mamba also needs registers. arXiv preprint arXiv:2405.14858

    2024 arXiv

  30. [38]

    Yang, N.; Wang, Y.; Liu, Z.; Li, M.; An, Y.; and Zhao, X. 2025. SMamba: Sparse Mamba for Event-based Object Detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 9229--9237

    2025

  31. [39]

    You, Z.; Cui, L.; Shen, Y.; Yang, K.; Lu, X.; Zheng, Y.; and Le, X. 2022. A unified model for multi-class anomaly detection. NeurIPS, 35: 4571--4584

    2022

  32. [40]

    Zavrtanik, V.; Kristan, M.; and Sko c aj, D. 2021. DRAEM-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 8330--8339

    2021

  33. [41]

    Zhang, J.; Chen, X.; Wang, Y.; Wang, C.; Liu, Y.; Li, X.; Yang, M.-H.; and Tao, D. 2023 a . Exploring plain vit reconstruction for multi-class unsupervised anomaly detection. arXiv e-prints

    2023

  34. [42]

    Zhang, X.; Li, S.; Li, X.; Huang, P.; Shan, J.; and Chen, T. 2023 b . DeSTSeg: Segmentation Guided Denoising Student-Teacher for Anomaly Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 3914--3923

    2023

  35. [43]

    Zhang, X.; Xu, M.; and Zhou, X. 2024. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 16699--16708

    2024

  36. [44]

    Zhu, L.; Liao, B.; Zhang, Q.; Wang, X.; Liu, W.; and Wang, X. 2024. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417

    2024 arXiv

  37. [45]

    Zou, Y.; Jeong, J.; Pemula, L.; Zhang, D.; and Dabeer, O. 2022. SPot -the- Difference Self - Supervised Pre -training for Anomaly Detection and Segmentation . ArXiv:2207.14315 [cs]

    2022 arXiv

Pith tools

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