Pith. sign in

REVIEW 3 major objections 6 minor 37 references

GCR: Geometry-Consistent Routing for Task-Agnostic Continual Anomaly Detection

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

Pith's one-line read Most continual-anomaly 'forgetting' is routing instability, not representation loss, this paper argues.

desk verdict GCR's routing rule is sensible and likely reproducible, but the paper's claim that continual forgetting is really routing instability is not supported by its own controlled ablation. read the letter →

arxiv 2601.01856 v3 pith:WW6Q2HYX submitted 2026-01-05 cs.CV

classification cs.CV
keywords continualanomalydetectiontask-agnosticinferenceroutingcatastrophicforgettingprototype-basedscoringmixture-of-expertsMVTecADVisA
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 targets task-agnostic continual anomaly detection, where product categories arrive one after another and the category identity of a test image is unknown. It claims that the steep performance drop seen when new categories are added—usually blamed on catastrophic forgetting of earlier knowledge—is largely caused by an unstable head-selection rule rather than degraded features. When a test image must be assigned to one of several independently built category-specific normality models, comparing their raw anomaly scores is unreliable because each head has its own score scale and tail behavior. The proposed method, GCR, avoids this by routing the image in a shared frozen patch-embedding space using accumulated nearest-prototype distance to per-category prototype banks, and only then computing the anomaly map within the selected head. On MVTec AD and VisA, this achieves near-zero forgetting while keeping detection and localization competitive, which the paper reads as evidence that the decision rule, not the features, was the primary bottleneck.

What carries the argument

The load-bearing object is the geometry-consistent routing rule: r_c(x) = (1/N) Σ_p min_k ||q_p(x) − μ_{c,k}||², with top-1 gating ĉ(x) = argmin_c r_c(x). Prototype banks {μ_{c,k}} are built per category by greedy k-center coreset selection over normal patch features extracted from a frozen Vision Transformer. Because routing and scoring share the same frozen embedding space and the same prototypes, the method does not require head-specific anomaly scores to be comparable across categories.

What would settle it

Construct a category set engineered to be visually confusable under the frozen backbone's features (e.g., overlapping product classes or a cross-dataset transfer) and measure GCR's routing accuracy at moderate prototype coverage. If routing accuracy falls materially below 100% and the overall AUROC drop matches the routing-error rate, the decision-rule explanation is supported; if misrouted samples still achieve near-oracle AUROC, the claim that routing instability dominates performance collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that a large portion of performance collapse in task-agnostic continual anomaly detection stems from cross-head decision instability in routing, not representational degradation. GCR separates routing from scoring: routing picks the category head by summing, over all patches, the squared distance from each patch embedding to the nearest prototype in that category's bank, all within a shared frozen patch-embedding space; scoring then uses a prototype-based energy only inside the routed head. In experiments, geometry-consistent routing achieves perfect routing accuracy on MVTec AD, routed performance matches oracle (category-known) performance once prototype covera

Load-bearing premise

The method assumes that accumulated nearest-prototype distance in the frozen patch-embedding space always identifies the correct category for both normal and anomalous images; the paper measures 100% routing accuracy on MVTec AD, but on harder or more visually similar category sets that perfect separation may break down, and with it the near-zero forgetting result.

Editorial extensions

If this is right

  • If the decision-rule explanation is correct, continual anomaly detection evaluations should report routing diagnostics (routing accuracy and conditional scores) alongside aggregate AUROC so that routing errors are not conflated with within-head scoring failures.
  • Near-zero forgetting can be achieved without end-to-end representation learning or rehearsal, as long as the routing criterion is shared and geometry-consistent across heads.
  • Routing by directly comparing anomaly scores from independently built heads is an ill-posed decision rule; methods that do this should degrade as head heterogeneity grows.
  • Once routing is fixed to the geometry-consistent rule, the exact within-head scoring form (energy vs. NLL, hard-min vs. LogSumExp) has only a minor effect on ranking metrics, so the observed gains are attributable to the routing rule itself.

Reading between the lines

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

  • If routing instability is the true bottleneck, a natural next step is to learn a routing metric in the shared embedding space (e.g., a small metric head) rather than learning per-category representations; this could transfer to visually similar category sets where pure nearest-prototype distance is less decisive.
  • The linear growth of per-category prototype banks with the number of categories is a scaling limit the paper acknowledges; a memory-sharing scheme that preserves the shared geometric criterion while capping memory is a testable extension.
  • The paper's own diagnostics imply that a benchmark with visually confusable categories would stress the perfect-routing assumption; such a stress test could reveal where routing accuracy degrades and forgetting reappears.
  • The core insight—that independent heads' score scales can destabilize cross-head decisions—likely generalizes beyond anomaly detection to any multi-head system where decisions are made by comparing head-specific scores.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes GCR, a training-free, prototype-based method for task-agnostic continual anomaly detection. It extracts multi-layer patch embeddings from a frozen CLIP ViT, builds per-category prototype banks by greedy k-center coreset selection, routes each test image by minimizing accumulated nearest-prototype distance to candidate banks (Eqs. 7–8), and then computes the anomaly map only within the routed head using a LogSumExp prototype distance (Eqs. 5–6). Experiments on MVTec AD and VisA report strong image-level AUROC and pixel-level p-AP, with numerically zero or near-zero forgetting. The paper's central interpretive claim is that a large portion of continual catastrophic forgetting in this setting is better understood as cross-head decision instability in routing rather than representational degradation.

Significance. If established, the decision-rule reframing would be a useful conceptual contribution: it would direct continual anomaly detection research toward routing and score-normalization rather than only representation learning. The method itself is attractive: it is simple, training-free, memory-bounded by the coreset size, and the authors provide routing diagnostics, a full K-sweep, efficiency measurements, and public code. However, the current evidence does not establish the central reframing claim, and the headline comparison to prior continual methods is explicitly uncontrolled. The paper is therefore best seen at this stage as a promising engineering proposal whose conceptual contribution needs stronger support.

major comments (3)
  1. [§4.3, Table 5] The key ablation does not support the paper's central reframing. Switching from score-based routing (arg min_c score_c(x)) to geometry-consistent routing (Eq. 8) improves routing accuracy from 0.937 to 1.000 and pixel-level p-AP from 0.380 to 0.508, but image-level AUROC remains exactly 98.08 in both conditions. Since the forgetting measure FM is defined on image-level AUROC (§4.2), this ablation provides no evidence that routing instability drives the detection/forgetting collapse. At most it supports a localization-specific effect. To support the abstract/conclusion claim, the authors need either a controlled condition in which image-level AUROC changes with routing, or an explicit argument for why the FM computed on AUROC is sensitive to routing when Table 5 shows it is not.
  2. [§3.2, §3.3, §4.2, Table 1] The near-zero FM of GCR is structurally guaranteed. The encoder is frozen (§3.2) and per-category prototype banks are fixed after coreset selection (§3.3); no shared parameters are updated, so there is no representation that can forget. The statement that GCR 'mitigates continual performance collapse' is therefore close to tautological. The stronger claim that 'continual failures previously attributed to catastrophic forgetting can, in part, be reframed as decision instability' requires a counterfactual experiment with a continually updated shared representation, or a controlled demonstration that an existing method's drop is caused by score-scale mismatch rather than representation drift. As written, the near-zero FM is an artifact of the frozen-head design, not evidence about the cause of forgetting in prior methods.
  3. [§4.1, Tables 1–4] The comparison to baselines is explicitly uncontrolled: backbones, input resolutions, and training configurations differ across methods. Given that, the large AUROC margins over UniAD and UCAD cannot be attributed to routing or to the proposed decision rule; they could reflect the stronger frozen CLIP backbone or higher-resolution features. The central contribution is the routing rule, so a controlled comparison is essential. I would expect at least a PatchCore-style baseline using the same frozen CLIP features, same prototype banks, and score-based routing (the condition in Table 5) reported on the full continual protocol, so that the route from routing accuracy to end-task performance can be assessed against matched baselines.
minor comments (6)
  1. [§3.5 vs. §4.1] Eq. (7) defines r_c(x) as a sum over N patches, while §4.1 writes r_c(x) with a 1/N normalization. The arg min is unaffected, but the notation should be consistent.
  2. [§3.3/§4.1, Table 7/Appendix C] The default prototype-bank size K is not stated in the main text, despite being a free parameter swept in Appendix C. Also, layer index 6 is selected on MVTec (Table 7); clarify whether the same layer was used for VisA and whether K was tuned per dataset. This matters for the claim of 'a single set of hyperparameters across all categories and both datasets.'
  3. [§4.2] The sentence 'we emphasize that GCR does not update a shared representation to prevent forgetting. Instead, heads remain category-specific, and FM primarily reflect' is grammatically incomplete; the intended conclusion appears to be cut off.
  4. [Tables 2 and 4] Table 4 reports p-AP but the FM in §4.2 is defined on image-level AUROC. Clarify which metric is used for the localization FM values, and make the bold-face conventions consistent (e.g., Table 2 bolds GCR on several rows where UniAD/UCAD are equal or better).
  5. [§5] The conclusion states that routing is performed under 'ℓ2-normalized patch features,' but ℓ2 normalization is not described in the method (§3.2) or in the routing definition (Eqs. 7–8). Either add the normalization to the method or remove it from the conclusion.
  6. [Throughout] There are several typos and spacing errors: 'na ıvely' in §5, 'N p=1' in Eq. (7) display, and 'AvgFM is lower is better' in table captions. These do not affect the technical content but should be cleaned up.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GCR's routing and scoring share prototypes by design, but no prediction reduces to a fitted input or to a self-citation chain.

full rationale

The paper's derivation chain is self-contained: frozen OpenCLIP features are extracted, per-category prototype banks are built via greedy k-center coreset selection, routing selects the head by accumulated nearest-prototype distance (Eqs. 7-8), and anomaly scoring is performed only within the routed head (Eqs. 5-6, 9-10). None of these steps defines a quantity in terms of the target it is claimed to predict. The routing rule is not fitted to test labels, and the core ablation (Table 5) holds the backbone and prototype banks fixed while varying only the routing rule, so the reported routing-accuracy gain is attributable to the routing design rather than to a circular fit. The near-zero FM is indeed a structurally guaranteed consequence of the frozen encoder and fixed per-category prototype banks—GCR does not update a shared representation, and FM is measured on image-level AUROC—but the paper does not disguise this as a prediction; it explicitly notes that heads remain category-specific. The broader claim that prior continual failures are, in part, decision-instability rather than representation-forgetting is an interpretation of the experiments, not a derivation from the equations. That interpretation may be under-supported, since the paper's own Table 5 shows image-level AUROC unchanged at 98.08 while routing accuracy improves, making the evidence at most localization-specific; however, under-support is a correctness/evidence concern, not circularity. There is no load-bearing self-citation: the cited UniAD/UCAD and PatchCore works are external, and no uniqueness theorem or prior result by the same authors is invoked to force the choice of routing rule. Therefore no specific circular reduction can be exhibited.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim depends on a small set of hyperparameters (K, layer index, temperature) and strong assumptions about the transferability of frozen CLIP features and the reliability of prototype-distance routing. There are no newly invented physical or mathematical entities. The layer index and K are effectively selected on the test set, which is a mild form of parameter fitting.

free parameters (4)
  • K (prototype bank size) = not specified in main tables; swept from 16 to 512 in Appendix C
    K controls the geometric coverage of each category's normal manifold. The main results do not state which K was used, and Appendix C suggests K=196 as a practical choice. This is a hyperparameter selected based on performance on the same benchmark.
  • layer_index (transformer block for features) = 6
    Chosen from Table 7 based on macro AUROC on MVTec AD (97.93). This selection uses test-set performance, which is a form of fitting to the benchmark.
  • M (number of patches used for routing approximation) = not specified
    Mentions approximating Eq. (7) with a random subset M << N, but the value is never given, affecting exact reproducibility.
  • temperature tau (LSE soft-min) = not specified
    In Eq. (5), tau controls softness. No default value is stated, though results may be insensitive as noted in Appendix B.
assumptions (3)
  • domain assumption Frozen OpenCLIP/CLIP ViT-B/16 features provide transferable patch representations for anomaly detection and routing across categories.
    The paper relies entirely on a frozen pre-trained encoder (Section 3.2) and does not learn representations. This is a standard assumption in feature-based AD, but it could fail for domains far from natural images.
  • domain assumption Euclidean nearest-prototype distance in the patch embedding space is a reliable indicator of category identity.
    The routing rule (Eq. 7-8) assumes that the category with the smallest accumulated distance to prototypes is the correct one. This is tested empirically (100% routing accuracy on MVTec AD) but is not guaranteed in general.
  • domain assumption The continual evaluation protocol (sequential category arrival, no task labels, evaluation on all seen categories) matches deployment conditions.
    The paper evaluates this protocol on MVTec AD and VisA, but real deployments may have different ordering, class imbalance, or domain shift. The protocol is standard in continual learning, but it is an assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GCR: Geometry-Consistent Routing for Task-Agnostic Continual Anomaly Detection." pith.science (2026). https://pith.science/paper/WW6Q2HYX

@misc{pith2026260101856,
  author       = {Pith},
  title        = {Pith review of: GCR: Geometry-Consistent Routing for Task-Agnostic Continual Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WW6Q2HYX}},
  note         = {Machine review of arXiv:2601.01856}
}
read the original abstract

Feature-based anomaly detection is widely adopted in industrial inspection due to the strong representational power of large pre-trained vision encoders. While most existing methods focus on improving within-category anomaly scoring, practical deployments increasingly require task-agnostic operation under continual category expansion, where the category identity is unknown at test time. In this setting, overall performance is often dominated by expert selection, namely routing an input to an appropriate normality model before any head-specific scoring is applied. However, routing rules that compare head-specific anomaly scores across independently constructed heads are unreliable in practice, as score distributions can differ substantially across categories in scale and tail behavior. We propose GCR, a lightweight mixture-of-experts framework for stabilizing task-agnostic continual anomaly detection through geometry-consistent routing. GCR routes each test image directly in a shared frozen patch-embedding space by minimizing an accumulated nearest-prototype distance to category-specific prototype banks, and then computes anomaly maps only within the routed expert using a standard prototype-based scoring rule. By separating cross-head decision making from within-head anomaly scoring, GCR avoids cross-head score comparability issues without requiring end-to-end representation learning. Experiments on MVTec AD and VisA show that geometry-consistent routing substantially improves routing stability and mitigates continual performance collapse, achieving near-zero forgetting while maintaining competitive detection and localization performance. These results indicate that many failures previously attributed to representation forgetting can instead be explained by decision-rule instability in cross-head routing. Code is available at https://github.com/jw-chae/GCR

Figures

Figures reproduced from arXiv: 2601.01856 by the authors.

Figure 1
Figure 1. Overview of GCR. (a) Training Phase) GCR extracts multi-layer patch features from a frozen Vision Transformer and concatenates them to form patch-level representations. For each category, a prototype memory bank is constructed via coreset selection to compactly represent nor￾mal patterns, without explicit parameter learning or anomaly scoring during training. (b) Inference Phase) Routing: Given a test image, the sys… view at source ↗
Figure 2
Figure 2. Qualitative anomaly localization results of [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 6 linked inside Pith

  1. [1]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Bae, J., Lee, J.H., Kim, S.: Pni: industrial anomaly detection using position and neighborhood information. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6373–6383 (2023)

  2. [2]

    International Journal of Computer Vision130(4), 947–969 (2022)

    Bergmann, P., Batzner, K., Fauser, M., Sattlegger, D., Steger, C.: Beyond dents and scratches: Logical constraints in unsupervised anomaly detection and localization. International Journal of Computer Vision130(4), 947–969 (2022)

  3. [3]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Bergmann, P., Fauser, M., Sattlegger, D., Steger, C.: Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9592–9600 (2019) 14

  4. [4]

    arXiv preprint arXiv:2112.09045 (2021)

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

  5. [5]

    arXiv preprint arXiv:1807.02011 (2018)

    Bergmann, P., L¨ owe, S., Fauser, M., Sattlegger, D., Steger, C.: Improving unsupervised defect segmentation by applying structural similarity to autoencoders. arXiv preprint arXiv:1807.02011 (2018)

  6. [6]

    In: Proceedings of the European conference on computer vision (ECCV)

    Chaudhry, A., Dokania, P.K., Ajanthan, T., Torr, P.H.: Riemannian walk for incremental learn- ing: Understanding forgetting and intransigence. In: Proceedings of the European conference on computer vision (ECCV). pp. 532–547 (2018)

  7. [7]

    In: International conference on pattern recognition

    Defard, T., Setkov, A., Loesch, A., Audigier, R.: Padim: a patch distribution modeling framework for anomaly detection and localization. In: International conference on pattern recognition. pp. 475–489. Springer (2021)

  8. [8]

    In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Deng, H., Li, X.: Anomaly detection via reverse distillation from one-class embedding. In: Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9737–9746 (2022)

Show all 37 references
  1. [9]

    In: European Conference on Computer Vision

    Gao, B.B.: Learning to detect multi-class anomalies with just one normal image prompt. In: European Conference on Computer Vision. pp. 454–470. Springer (2024)

  2. [10]

    In: Proceedings of the AAAI conference on artificial intelligence

    Gu, Z., Zhu, B., Zhu, G., Chen, Y., Tang, M., Wang, J.: Anomalygpt: Detecting industrial anomalies using large vision-language models. In: Proceedings of the AAAI conference on artificial intelligence. vol. 38, pp. 1932–1940 (2024)

  3. [11]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Gu, Z., Liu, L., Chen, X., Yi, R., Zhang, J., Wang, Y., Wang, C., Shu, A., Jiang, G., Ma, L.: Remembering normality: Memory-guided knowledge distillation for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 16401–...

  4. [12]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Gudovskiy, D., Ishizaka, S., Kozuka, K.: Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 98–107 (2022)

  5. [13]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Guo, J., Lu, S., Zhang, W., Chen, F., Li, H., Liao, H.: Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 20405–20415 (2025)

  6. [14]

    In: Proceedings of the AAAI conference on artificial intelligence

    He, H., Zhang, J., Chen, H., Chen, X., Li, Z., Chen, X., Wang, Y., Wang, C., Xie, L.: A diffusion- based framework for multi-class anomaly detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 38, pp. 8472–8480 (2024)

  7. [15]

    In: European Conference on Computer Vision

    He, L., Jiang, Z., Peng, J., Zhu, W., Liu, L., Du, Q., Hu, X., Chi, M., Wang, Y., Wang, C.: Learn- ing unified reference representation for unsupervised multi-class anomaly detection. In: European Conference on Computer Vision. pp. 216–232. Springer (2024)

  8. [16]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

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

  9. [17]

    IEEE Access10, 78446–78454 (2022)

    Lee, S., Lee, S., Song, B.C.: Cfa: Coupled-hypersphere-based feature adaptation for target- oriented anomaly localization. IEEE Access10, 78446–78454 (2022)

  10. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Li, X., Zhang, Z., Tan, X., Chen, C., Qu, Y., Xie, Y., Ma, L.: 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. pp. 16838–16848 (2024)

  11. [19]

    In: Proceedings of the AAAI conference on artificial intelligence

    Liu, J., Wu, K., Nie, Q., Chen, Y., Gao, B.B., Liu, Y., Wang, J., Wang, C., Zheng, F.: Unsu- pervised continual anomaly detection with contrastively-learned prompt. In: Proceedings of the AAAI conference on artificial intelligence. vol. 38, pp. 3639–3647 (2024) 15

  12. [20]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, Z., Zhou, Y., Xu, Y., Wang, Z.: Simplenet: A simple network for image anomaly detection and localization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20402–20411 (2023)

  13. [21]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Roth, K., Pemula, L., Zepeda, J., Sch¨ olkopf, B., Brox, T., Gehler, P.: Towards total recall in industrial anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14318–14328 (2022)

  14. [22]

    In: Proceedings of the IEEE/CVF winter conference on applications of computer vision

    Rudolph, M., Wandt, B., Rosenhahn, B.: Same same but differnet: Semi-supervised defect detec- tion with normalizing flows. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 1907–1916 (2021)

  15. [23]

    arXiv preprint arXiv:2111.02114 (2021)

    Schuhmann, C., Vencu, R., Beaumont, R., Kaczmarczyk, R., Mullis, C., Katta, A., Coombes, T., Jitsev, J., Komatsuzaki, A.: Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114 (2021)

  16. [24]

    arXiv preprint arXiv:2110.03396 (2021)

    Song, J., Kong, K., Park, Y.I., Kim, S.G., Kang, S.J.: Anoseg: Anomaly segmentation network using self-supervised learning. arXiv preprint arXiv:2110.03396 (2021)

  17. [25]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Sun, H., Cao, Y., Dong, H., Fink, O.: Unseen visual anomaly generation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 25508–25517 (2025)

  18. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Wang, C., Zhu, W., Gao, B.B., Gan, Z., Zhang, J., Gu, Z., Qian, S., Chen, M., Ma, L.: Real- iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. ...

  19. [27]

    arXiv preprint arXiv:2103.04257 (2021)

    Wang, G., Han, S., Ding, E., Huang, D.: Student-teacher feature pyramid matching for anomaly detection. arXiv preprint arXiv:2103.04257 (2021)

  20. [28]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Wyatt, J., Leach, A., Schmon, S.M., Willcocks, C.G.: Anoddpm: Anomaly detection with de- noising diffusion probabilistic models using simplex noise. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 650–656 (2022)

  21. [29]

    In: Proceedings of the Asian conference on computer vision (2020)

    Yi, J., Yoon, S.: Patch svdd: Patch-level svdd for anomaly detection and segmentation. In: Proceedings of the Asian conference on computer vision (2020)

  22. [30]

    Advances in Neural Information Processing Systems35, 4571–4584 (2022)

    You, Z., Cui, L., Shen, Y., Yang, K., Lu, X., Zheng, Y., Le, X.: A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems35, 4571–4584 (2022)

  23. [31]

    arXiv preprint arXiv:2111.07677 (2021)

    Yu, J., Zheng, Y., Wang, X., Li, W., Wu, Y., Zhao, R., Wu, L.: Fastflow: Unsupervised anomaly detection and localization via 2d normalizing flows. arXiv preprint arXiv:2111.07677 (2021)

  24. [32]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Zavrtanik, V., Kristan, M., Skoˇ caj, D.: Draem-a discriminatively trained reconstruction embed- ding for surface anomaly detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8330–8339 (2021)

  25. [33]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    Zhang, H., Wang, Z., Zeng, D., Wu, Z., Jiang, Y.G.: Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

  26. [34]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zhang, X., Li, N., Li, J., Dai, T., Jiang, Y., Xia, S.T.: Unsupervised surface anomaly detection with diffusion probabilistic model. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6782–6791 (2023)

  27. [35]

    arXiv preprint arXiv:2310.18961 (2023)

    Zhou, Q., Pang, G., Tian, Y., He, S., Chen, J.: Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2310.18961 (2023)

  28. [36]

    Neurocomputing453, 131–140 (2021)

    Zhou, Y., Liang, X., Zhang, W., Zhang, L., Song, X.: Vae-based deep svdd for anomaly detection. Neurocomputing453, 131–140 (2021)

  29. [37]

    In: European conference on computer vision

    Zou, Y., Jeong, J., Pemula, L., Zhang, D., Dabeer, O.: Spot-the-difference self-supervised pre- training for anomaly detection and segmentation. In: European conference on computer vision. pp. 392–408. Springer (2022) 16 A Notation and Dimensional Consistency For clarity, we s...

Pith tools

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