Pith. sign in

REVIEW 4 major objections 5 minor 56 references

ArcAD: Anomaly-Rectified Calibration for Cold-Start Supervised Anomaly Detection

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a plug-and-play calibration layer built from balanced hyperspherical clustering and defect-guided pseudo-anomaly synthesis makes reconstruction-based anomaly detectors substantially more accurate in cold-start settings

desk verdict ArcAD's core calibration idea is new and the additive gains over reconstruction baselines largely hold, but the SOTA claim is undermined by single-run, under-tuned supervised baselines and small deltas on the easy datasets. read the letter →

arxiv 2607.02252 v2 pith:NUMGXKOW submitted 2026-07-02 cs.CV

Ningning Han , Lei Fan , Jia Guo , Yunkang Cao , Xiu Su , Feng Cao , Donglin Di , Tonghua Su This is my paper
classification cs.CV
keywords cold-startanomalydetectionindustrialhyperspherelearningSinkhornoptimaltransportprototypeclusteringpseudo-anomalysynthesiscontrastivecalibrationreconstruction-based
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

ArcAD is a training-time layer for reconstruction-based industrial anomaly detectors under cold-start scarcity: few normal images, very few defects. Its claim is that by projecting patch features onto a unit hypersphere, partitioning them into uniformly balanced clusters via a Sinkhorn optimal-transport assignment, and then synthesizing boundary pseudo-anomalies that real defects push away from the normal prototypes, a compact and discriminative normal boundary emerges. The authors report that attaching ArcAD to three published reconstruction detectors improves image-level AUROC on four industrial datasets, e.g. raising one baseline from 88.8 to 92.5 on Real-IAD multi-class and from 90.3 to 93.3 on MANTA, while leaving the inference-time anomaly score unchanged. If right, the framework would let factories exploit the handful of defects collected during line ramp-up without retraining the detector or slowing it down.

What carries the argument

The load-bearing object is a unit hypersphere latent space together with a balanced assignment matrix. SPM computes a soft feature-to-prototype assignment by solving an entropically regularized optimal transport problem with Sinkhorn–Knopp iterations, enforcing that each of K prototypes receives equal total mass; this yields compact, uniformly spread normal clusters even when normal data are scarce. DGC then generates candidate perturbations of normal features on the sphere, filters them by their maximum similarity to the learned prototypes to keep the hardest pseudo-anomalies, and applies a contrastive calibration loss that repels real and synthetic anomalies from the normal prototypes whil

What would settle it

Re-run the authors' cold-start protocol on Real-IAD with every baseline tuned by a neutral party, training anomalies drawn from a split held out of the test set, no pixel masks given to ArcAD (image-level labels only), and 95% confidence intervals over several seeds; if ArcAD's margin over the strongest reconstruction baseline collapses below about 1 I-AUROC or reverses, the central claim is false.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that data scarcity in cold-start anomaly detection should be treated as a geometry problem rather than a data-augmentation problem. Normal patch features live on the unit hypersphere and are organized into K clusters with an equipartition constraint (SPM), so no single dense region absorbs all normal mass; then a prototype-restricted synthesis step generates pseudo-anomalies near the boundary, and a contrastive loss (DGC) pushes real and synthetic anomalies away from the nearest normal prototypes. The combined push–pull calibration makes the reconstruction-based latent space compact and discriminative, and the ablation attributes +3.7 I-AUROC on Rea

Load-bearing premise

The argument stands or falls on whether the cold-start benchmark is a fair instrument: if the reimplemented baselines are under-tuned, or ArcAD's extra pixel-mask supervision explains its edge, the headline claim is a protocol artifact.

Editorial extensions

If this is right

  • If the claim holds, a factory can add ArcAD to an existing reconstruction detector and use the few defects logged during ramp-up to sharpen detection, with no extra inference cost.
  • Reported gains are largest on broad multi-class datasets: +3.7 I-AUROC on Real-IAD and +3.0 on MANTA over the strongest unsupervised baseline, and up to +11.2 for a weaker reconstruction baseline.
  • The improvements persist across three different reconstruction backbones, suggesting the calibration is not tied to one architecture.
  • ArcAD still helps when only 3% of training data are anomalies, so the method tolerates even more extreme scarcity.
  • Single-class performance also improves, reaching 100.0 I-AUROC on MVTec-AD and gaining +3.4 on Real-IAD.

Reading between the lines

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

  • Editorial inference: because ArcAD only regularizes the training latent space, its push–pull mechanics should transfer to other one-class or reconstruction pipelines beyond the three tested — but that transfer is untested in the paper.
  • Editorial inference: the cold-start protocol draws the few training anomalies from the original test set, meaning the model has seen some test anomalies in training; a stricter held-out anomaly split would test whether the reported gains survive on truly unseen defects.
  • Editorial inference: DGC uses pixel-level ground-truth masks to select anomaly patches, supervision that image-level-labeled baselines do not receive; if ArcAD is restricted to image-level labels, its margin over those baselines may shrink.
  • Editorial inference: the Sinkhorn equipartition idea could be reused for partial-label or noisy-label anomaly settings, where the balanced assignment would resist majority-class domination.
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 ArcAD, a plug-and-play training-time calibration layer for reconstruction-based industrial anomaly detectors under a cold-start regime with few normal samples and rare anomalies. It maps bottleneck patch features to the unit hypersphere and organizes them into compact, uniformly distributed clusters via Sinkhorn-based prototype modeling (SPM). A defect-guided calibration module (DGC) synthesizes pseudo-anomalies by perturbing normal features and filtering them with prototype similarity (Eqs. 5–6), then uses a contrastive loss (Eq. 8) plus a focal-loss discriminator (Eq. 9) to push real and synthetic anomalies away from normal prototypes. Inference uses the unchanged reconstruction score. The authors report consistent gains over unsupervised and supervised baselines on MVTec-AD, VisA, Real-IAD, and MANTA, with the largest improvements on Real-IAD (+3.7 I-AUROC over Dinomaly) and MANTA (+11.2 for RD4AD).

Significance. If the results are reproducible, ArcAD is a practically useful contribution: it improves several diverse reconstruction baselines without changing inference-time scoring, and the ablation in Table 4 gives a plausible role for each component. The manuscript is clearly written, and the internal arithmetic is consistent: the +3.7 Real-IAD delta and the Fig. 4 gains agree with Tables 1 and 3. The main value is the demonstration that a light clustering-plus-synthesis regularizer can alleviate cold-start data scarcity. However, the claimed 'significantly outperforms state-of-the-art' result is not yet supported by the evidence: the evaluation is single-run, the baselines are reimplemented without released configs, and some baseline results are close to chance. These issues are load-bearing for the headline claim, so the manuscript requires major revision.

major comments (4)
  1. [§4.1, Tables 1–2] All results are reported as single numbers with no seeds or error bars. The headline improvements on MVTec-AD and VisA are small (multi-class I-AUROC +0.1 and +0.5; single-class +0.5 and +0.2), well within typical rerun variance for such benchmarks. The abstract's 'significantly outperforms' is therefore unsupported on these datasets. Please report mean and standard deviation over at least 3–5 seeds, and where possible a paired test for the ArcAD-vs-baseline delta. At minimum, state the number of runs and the observed run-to-run variability.
  2. [§4.1, Table 1] The supervised baselines are all reimplemented by the authors under the cold-start protocol, but no configurations or training details are released. Several results are red flags: DevNet scores 50.0 I-AUROC on Real-IAD multi-class (chance), SDNet P-F1-max is 0.9–4.2 across datasets, and SDNet degrades from 70.1 to 59.7 I-AUROC on MVTec-AD when its backbone is upgraded to Dinov2, contrary to the expected feature-quality trend. If these baselines are under-tuned, the reported +14.5 to +19.6 point gains over supervised methods are protocol artifacts. Please release the reimplementation configs, use official code where available, and verify that DevNet and SDNet are not broken under the cold-start protocol.
  3. [Eq. (7), Table 1] ArcAD is trained with pixel-level ground-truth masks to isolate anomalous patches, whereas DRA and DevNet use image-level labels. This gives ArcAD an extra source of supervision that is not controlled in the comparison. Please compare against an image-level-label variant of ArcAD (e.g., using the whole anomaly image or only image-level labels), or include mask-supervised supervised baselines under the same protocol, so that the comparison is equitable.
  4. [§4.3, Fig. 5] The number of prototypes K=500 is selected based on I-AUROC on the Real-IAD test set, and the same test set is then used for the final reported numbers. This is test-set selection and can inflate the reported gain on Real-IAD. Use a validation split for K and report the model at the selected K, or show that the results are stable across K. In addition, the noise scale σ, temperatures ε and τ, and the three loss weights λ1–λ3 are fixed (λ1=λ2=λ3=0.1) without sensitivity analysis; please provide a small sensitivity study for the most influential of these hyperparameters.
minor comments (5)
  1. [Fig. 4] The gain labels for small deltas (e.g., +0.1, +0.2) are very hard to read. Consider using a table or larger fonts, or only annotating deltas above a threshold.
  2. [Table 3] The header 'Official' vs 'Cold Start' is confusing: 'Official' appears to be the standard-setting result quoted from prior work, not a result of this paper. Please clarify the column definitions and the subscript convention.
  3. [Eq. (9)] The binary label y is described as 0 for normal and 1 for abnormal, but the loss expression does not make explicit how the focal loss weights positives versus negatives. A short clarification would help.
  4. [§4.1] State explicitly whether the anomaly images sampled from the original test set for training are excluded from the cold-start test set. The description 'test set ... contains reduced anomalies' is ambiguous; the per-dataset counts in Fig. 3 help but the exact split rule should be stated in one sentence.
  5. [§4.3, Table 6] Minor wording issues: 'utilizes all candidate anomalies is sub-optimal' should read 'utilizing all candidate anomalies is sub-optimal.' Also, the sentence beginning 'Notably, compared to the strategy without filter' is missing a subject.

Circularity Check

1 steps flagged · score 5.0 of 10

Real-IAD headline gain is the K-selection optimum on the same test set; ArcAD's inference-time score is otherwise independent of the fitted prototypes.

  1. fitted input called prediction [Sec. 4.3, Fig. 5 (K ablation); Table 1; Intro '+3.7%' Real-IAD claim]
    "The model reaches its highest I-AUROC (92.5%) at K= 500, with P-AUROC staying relatively stable above 99.0%. Although the highest P-F1-max (50.2%) occurs at K= 800, we ultimately select K= 500 as it provides the best balance between image-level and pixel-level detection performance."

    The prototype count K is chosen by running ArcAD on the Real-IAD multi-class test set and reading the I-AUROC curve (Fig. 5). Table 1 then reports ArcAD on that same Real-IAD multi-class test set as 92.5 I-AUROC, exactly the value used to select K; the abstract/introduction summarizes the same number as the +3.7% gain over Dinomaly. Thus the headline Real-IAD result is not an independent hold-out evaluation: it is the optimum of the tuning curve used to set K. Other datasets are evaluated with the same K but were not used in the sweep, so the circularity is partial.

full rationale

The proposed framework itself is not definitionally circular: at inference the anomaly score is the unmodified reconstruction discrepancy, and SPM/DGC only shape the training-time latent space. No theoretical prediction reduces to a fitted value by construction. The concrete circular step is experimental: K is tuned on the Real-IAD test set and the same test set yields the flagship 92.5 I-AUROC / +3.7 result. The paper also uses self-cited anchors (Dinomaly [14], standard-setting reference [12], MANTA [8]) and states all baselines were reimplemented by the authors, but these are reproducibility/comparison-validity concerns rather than load-bearing self-citation circularity, so they do not independently raise the score. Absence of error bars and implausible baseline numbers (DevNet 50.0 chance, SDNet Dinov2 regression) are correctness risks, not circularity. Because the central SOTA claim has independent support on other datasets (e.g., MANTA +3.0) and the training objectives do not enter inference scoring, the circularity burden is moderate (5/10) rather than severe.

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

The central claim depends primarily on trained quantities: K=500 prototypes (test-tuned), synthesis noise σ (unreported), Sinkhorn temperature ε and vMF temperature τ (unreported), per-batch Sinkhorn mass (batch size unreported), and evenly set loss weights (λ=0.1). Everything else the paper gains, it pulls from four families of prior work: reconstruction backbones [5,13,14], hypersphere/vMF geometry [4,29,31], Sinkhorn balanced assignment [2], and perturbation-based pseudo-anomalies [21,26,51]. No free constant is being 'predicted': the output metric (reconstruction AUROC) is not a function of the fitted prototypes, which is why circularity burden stays moderate despite the tuning-on-test issue. No new physical or conceptual entities are postulated; the hypersphere prototypes are learned parameters and the pseudo-anomalies are synthesized training samples, so the invented-entities ledger is empty.

free parameters (6)
  • Prototype count K = 500
    Selected by ablation on Real-IAD multi-class test metrics (Fig. 5); controls manifold coverage in SPM and the distance reference in DGC; used for all main results.
  • Noise scale σ for pseudo-anomaly synthesis (Eq. 5)
    Determines how far synthesized anomalies move on the hypersphere; never reported in the paper, blocking exact reproduction.
  • Sinkhorn temperature ε (Eq. 3)
    Entropic-OT regularization temperature controlling assignment sharpness; unspecified.
  • vMF temperature τ = 1/κ (Eq. 4)
    Logit temperature in the prototype cross-entropy; unspecified.
  • Loss weights λ1=λ2=λ3 = 0.1
    Set uniformly by hand in §4.1 with no sensitivity analysis.
  • Batch size (N_batch)
    Not reported; per-batch Sinkhorn assignment and prototype EMA updates make results batch-size dependent.
assumptions (6)
  • domain assumption Normalized patch features follow a von Mises-Fisher distribution with a single concentration parameter per prototype
    Sec. 3.2/Eq. 2 justifies L_spm; prior work [4,38] establishes Gaussian-like cluster structure, but vMF fidelity is asserted, not tested on cold-start features.
  • domain assumption Balanced equipartition of features across prototypes yields better normal-manifold coverage than unconstrained clustering
    Sec. 3.2 ('strictly enforces uniformity to ensure no single prototype dominates'); imported from SSL practice, no AD-specific evidence given.
  • ad hoc to paper Isotropic Gaussian perturbation followed by ℓ2 renormalization of a normal feature produces viable pseudo-anomalies near the normal boundary
    Eq. 5; noise explores all sphere directions equally, yet real defects occupy specific directions; no argument links isotropic noise to defect directionality.
  • domain assumption The min-max candidate (Eq. 6) is the right pseudo-anomaly: the candidate farthest from normal prototypes is the hardest and most useful
    Validated only by ablation (Table 6: w/o Filter 90.2 vs Ours 92.5), not derived.
  • domain assumption Training-time boundary calibration transfers to reconstruction-error scoring at inference
    Sec. 3.4 keeps the standard reconstruction score; the coupling between L_dgc/L_cls and reconstruction error is never characterized.
  • standard math Sinkhorn-Knopp iteration converges to the entropically regularized OT solution
    Eq. 3 per Cuturi [2].

how reviews work

0 comments
Cite this review

Pith. "Pith review of ArcAD: Anomaly-Rectified Calibration for Cold-Start Supervised Anomaly Detection." pith.science (2026). https://pith.science/paper/NUMGXKOW

@misc{pith2026260702252,
  author       = {Pith},
  title        = {Pith review of: ArcAD: Anomaly-Rectified Calibration for Cold-Start Supervised Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NUMGXKOW}},
  note         = {Machine review of arXiv:2607.02252}
}
read the original abstract

The deployment of Industrial Anomaly Detection (IAD) in real-world manufacturing frequently encounters a challenging cold-start bottleneck, in which limited normal samples fail to represent the full normal distribution and only a few anomalies are available. Under such a regime, existing methods struggle to form compact normal boundaries and fail to effectively exploit supervised signals from rare defects. To address this challenge, we propose Anomaly-Rectified Cold-start AD (ArcAD), a plug-and-play calibration framework for reconstruction-based IAD baselines. ArcAD follows a push-pull learning paradigm to construct a compact and discriminative normal boundary under data scarcity. On the one hand, ArcAD projects limited normal samples onto a hypersphere and pulls them into multiple compact clusters to maximize coverage of the normal manifold. On the other hand, it synthesizes pseudo-anomalies on the hypersphere and leverages real anomalies to push the boundary inward and sharpen anomaly discrimination. Extensive experiments on MVTec-AD, VisA, Real-IAD, and MANTA demonstrate that ArcAD significantly outperforms state-of-the-art supervised and unsupervised methods in both single-class and multi-class settings under cold-start conditions. Code is available at: https://github.com/LGC-AD/ArcAD.

Figures

Figures reproduced from arXiv: 2607.02252 by the authors.

Figure 1
Figure 1. Left: Decision boundaries in cold-start scenarios. Unsupervised methods (a) yield loose manifolds and supervised methods (b) overfit to known anomalies, whereas ArcAD (c) forms compact and unified boundaries. Right: Paradigm comparison on data efficiency, anomaly utilization, and generalization. frequently encounters a severe cold-start bottleneck. Unlike standard unsuper￾vised settings [12, 41] that assume abundant… view at source ↗
Figure 2
Figure 2. Overview of the proposed ArcAD. (a) Integration of the ArcAD framework into a standard reconstruction-based architecture. (b) ArcAD explicitly calibrates the latent feature distribution through Sinkhorn-based Prototype Modeling (SPM) and Defect-Guided Calibration (DGC). 3.2 Sinkhorn-based Prototype Modeling (SPM) vMF Modeling. Compared to the unbounded Euclidean space, the hypersphere provides a compact manifold [27… view at source ↗
Figure 3
Figure 3. Statistics of standard and cold-start setting datasets (see details in Supp). In our cold-start setting, 30% of the standard training set is utilized as nor￾mal training data, which is then combined with a subset of anomalies sampled from the original test set. Typically, anomalies account for 10% of this newly constructed training set. Due to the severe scarcity of anomalous samples, we set this ratio to 5% for Vis… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Performance of baseline methods (Dinomaly, ReContrast, and RD4AD) equipped with the ArcAD across four datasets under the multi-class setting. Single-class Performance. We evaluated ArcAD under the single-class setting across four datasets, as summarized in [PITH_FULL_…
Figure 6
Figure 6. Figure 6: Left: Qualitative results of anomaly localization for multi-class anomaly de￾tection. Test images with anomalous regions are highlighted in red contours. Right: Feature visualization of ArcAD before and after optimization on the MANTA dataset. egy yields hard anomalies…
Figure 5
Figure 5. Figure 5: Ablation of the number of proto￾types K in the SPM. Number of Prototypes K. To evalu￾ate the impact of the prototype quan￾tity in the SPM module, we con￾ducted an ablation study by varying the number of prototypes K across {100, 300, 500, 800, 1000}. The model reaches …

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 4 linked inside Pith

  1. [1]

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

    Bergmann, P., Fauser, M., Sattlegger, D., Steger, C.: Mvtec ad–a comprehen- sive real-world dataset for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 9592– 9600 (2019)

    2019

  2. [2]

    Ad- vances in neural information processing systems26(2013)

    Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Ad- vances in neural information processing systems26(2013)

    2013

  3. [3]

    In: 34th Conference on Uncertainty in Artificial Intelligence 2018, UAI 2018

    Davidson, T.R., Falorsi, L., et al.: Hyperspherical variational auto-encoders. In: 34th Conference on Uncertainty in Artificial Intelligence 2018, UAI 2018. pp. 856–

    2018

  4. [4]

    In: International conference on pattern recognition

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

    2021

  5. [5]

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

    Deng, H., Li, X.: Anomaly detection via reverse distillation from one-class em- bedding. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 9737–9746 (2022)

    2022

  6. [6]

    In: Proceedings of the IEEE/CVF conference on Com- puter Vision and Pattern Recognition

    Ding, C., Pang, G., Shen, C.: Catching both gray and black swans: Open-set super- vised anomaly detection. In: Proceedings of the IEEE/CVF conference on Com- puter Vision and Pattern Recognition. pp. 7388–7398 (2022) 16 Han et al

    2022

  7. [7]

    IEEE Trans- actions on Industrial Informatics21(5), 3627–3635 (2025)

    Fan, L., Fan, D., Ding, Y., Wu, Y., Di, D., Pagnucco, M., Song, Y.: Grainbrain: Multiview identification and stratification of defective grain kernels. IEEE Trans- actions on Industrial Informatics21(5), 3627–3635 (2025)

    2025

  8. [8]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Fan, L., Fan, D., Hu, Z., Ding, Y., Di, D., Yi, K., Pagnucco, M., Song, Y.: Manta: A large-scale multi-view and visual-text anomaly detection dataset for tiny objects. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 25518–25527 (2025)

    2025

Show all 56 references
  1. [9]

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

    Fan, L., Huang, J., Di, D., Su, A., Song, T., Pagnucco, M., Song, Y.: Salvaging the overlooked: Leveraging class-aware contrastive learning for multi-class anomaly de- tection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 21419–21428 (2025)

    2025

  2. [10]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Fu, L., Huang, S., Lai, Y., Liao, T., Zhang, C., Chen, C.: Beyond federated proto- typelearning:Learnablesemanticanchorswithhypersphericalcontrastfordomain- skewed data. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 16648–16656 (2025)

    2025

  3. [11]

    In: International Confer- ence on Machine Learning

    Gopal, S., Yang, Y.: Von mises-fisher clustering models. In: International Confer- ence on Machine Learning. pp. 154–162. PMLR (2014)

    2014

  4. [12]

    arXiv preprint arXiv:2510.17611 (2025)

    Guo,J.,Lu,S.,Fan,L.,Li,Z.,Di,D.,Song,Y.,Zhang,W.,Zhu,W.,Yan,H.,Chen, F., et al.: One dinomaly2 detect them all: A unified framework for full-spectrum unsupervised anomaly detection. arXiv preprint arXiv:2510.17611 (2025)

    2025

  5. [13]

    Advances in Neural Information Processing Systems36, 10721–10740 (2023)

    Guo, J., Lu, S., Jia, L., Zhang, W., Li, H.: Recontrast: Domain-specific anomaly de- tection via contrastive reconstruction. Advances in Neural Information Processing Systems36, 10721–10740 (2023)

    2023

  6. [14]

    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)

    2025

  7. [15]

    Advances in Neural Information Processing Systems37, 71162– 71187 (2024)

    He, H., Bai, Y., Zhang, J., He, Q., Chen, H., Gan, Z., Wang, C., Li, X., Tian, G., Xie, L.: Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems37, 71162– 71187 (2024)

    2024

  8. [16]

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

    He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 9729–9738 (2020)

    2020

  9. [17]

    In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision

    Hwang, J.J., Yu, S.X., et al.: Segsort: Segmentation by discriminative sorting of segments. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision. pp. 7334–7344 (2019)

    2019

  10. [18]

    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)

    2023

  11. [19]

    arXiv preprint arXiv:2509.24335 (2025)

    Ke, G., Xue, H.: Hyperspherical latents improve continuous-token autoregressive generation. arXiv preprint arXiv:2509.24335 (2025)

    2025

  12. [20]

    In: Computer Vision and Pattern Recognition

    Kobayashi, T.: T-vmf similarity for regularizing intra-class feature distribution. In: Computer Vision and Pattern Recognition. pp. 6616–6625 (2021)

    2021

  13. [21]

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

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

    2021

  14. [22]

    In: The Thirteenth International Conference on Learning Representations ArcAD 17

    Li, C., Shi, Y., Hu, J., Zhu, X.X., Mou, L.: Scale-aware contrastive reverse distilla- tion for unsupervised medical anomaly detection. In: The Thirteenth International Conference on Learning Representations ArcAD 17

  15. [23]

    In: European Conference on Computer Vision

    Li, J.L., Chang, M.C., Chen, W.C.: Learning multi-manifold embedding for out-of- distribution detection. In: European Conference on Computer Vision. pp. 417–433. Springer (2024)

    2024

  16. [24]

    In: Proceedings of the IEEE International Conference on Computer Vision

    Lin, T.Y., Goyal, P., Girshick, R., He, K., Dollár, P.: Focal loss for dense object de- tection. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 2980–2988 (2017)

    2017

  17. [25]

    arXiv preprint arXiv:2511.20169 (2025)

    Ling, H., Guo, J., Tao, Z., Cao, Y., Di, D., Xu, H., Su, X., Song, Y., Fan, L.: Adnet: A large-scale and extensible multi-domain benchmark for anomaly detection across 380 real-world categories. arXiv preprint arXiv:2511.20169 (2025)

    2025

  18. [26]

    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)

    2023

  19. [27]

    In: The Thirteenth International Conference on Learning Representations

    Loshchilov, I., Hsieh, C.P., Sun, S., Ginsburg, B.: ngpt: Normalized transformer with representation learning on the hypersphere. In: The Thirteenth International Conference on Learning Representations

  20. [28]

    arXiv preprint arXiv:2402.02653 (2024)

    Lu, H., Gong, D., et al.: Learning with mixture of prototypes for out-of-distribution detection. arXiv preprint arXiv:2402.02653 (2024)

    2024 arXiv

  21. [29]

    John Wiley & Sons (2009)

    Mardia, K.V., Jupp, P.E.: Directional statistics. John Wiley & Sons (2009)

    2009

  22. [30]

    In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision

    McIntosh, D., Albu, A.B.: Inter-realization channels: Unsupervised anomaly de- tection beyond one-class classification. In: Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision. pp. 6285–6295 (2023)

    2023

  23. [31]

    Ming, Y., Sun, Y., et al.: How to exploit hyperspherical embeddings for out-of- distribution detection? arXiv preprint arXiv:2203.04450 (2022)

    2022 arXiv

  24. [32]

    IEEE Signal processing magazine13(6), 47–60 (1996)

    Moon, T.K.: The expectation-maximization algorithm. IEEE Signal processing magazine13(6), 47–60 (1996)

    1996

  25. [33]

    Transactions on Machine Learning Research Journal (2024)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research Journal (2024)

    2024

  26. [34]

    arXiv preprint arXiv:2108.00462 (2021)

    Pang, G., Ding, C., Shen, C., Hengel, A.v.d.: Explainable deep few-shot anomaly detection with deviation networks. arXiv preprint arXiv:2108.00462 (2021)

    2021 arXiv

  27. [35]

    In: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining

    Pang, G., Shen, C., Van Den Hengel, A.: Deep anomaly detection with deviation networks. In: Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. pp. 353–362 (2019)

    2019

  28. [36]

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

    Park, H., Noh, J., Ham, B.: Learning memory-guided normality for anomaly de- tection. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 14372–14381 (2020)

    2020

  29. [37]

    Journal of Machine Learning Research17(200), 1–51 (2016)

    Qin, X., Cunningham, P., et al.: Online trans-dimensional von mises-fisher mix- ture models for user profiles. Journal of Machine Learning Research17(200), 1–51 (2016)

    2016

  30. [38]

    IEEE Trans- actions on Instrumentation and Measurement70, 1–13 (2021)

    Rippel, O., Mertens, P., König, E., Merhof, D.: Gaussian anomaly detection by modeling the distribution of normal data in pretrained deep features. IEEE Trans- actions on Instrumentation and Measurement70, 1–13 (2021)

    2021

  31. [39]

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

    Roth, K., Pemula, L., Zepeda, J., Schölkopf, 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)

    2022

  32. [40]

    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 detection with normalizing flows. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 1907–1916 (2021) 18 Han et al

    1907

  33. [41]

    Proceedings of the IEEE109(5), 756–795 (2021)

    Ruff, L., Kauffmann, J.R., Vandermeulen, R.A., Montavon, G., Samek, W., Kloft, M., Dietterich, T.G., Müller, K.R.: A unifying review of deep and shallow anomaly detection. Proceedings of the IEEE109(5), 756–795 (2021)

    2021

  34. [42]

    Journal of Intelligent Manufacturing31(3), 759–776 (2020)

    Tabernik, D., Šela, S., Skvarč, J., Skočaj, D.: Segmentation-based deep-learning approach for surface-defect detection. Journal of Intelligent Manufacturing31(3), 759–776 (2020)

    2020

  35. [43]

    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. 2...

    2024

  36. [44]

    In: Proceed- ings of the IEEE/CVF conference on Computer Vision and Pattern Recognition

    Wang, F., Zhang, T., Wang, Y., Qiu, Y., Liu, X., Guo, X., Cui, Z.: Distribution pro- totype diffusion learning for open-set supervised anomaly detection. In: Proceed- ings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 20416–20426 (2025)

    2025

  37. [45]

    In: International conference on ma- chine learning

    Wang, T., Isola, P.: Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In: International conference on ma- chine learning. pp. 9929–9939. PMLR (2020)

    2020

  38. [46]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Wei, S., Jiang, J., Xu, X.: Uninet: A contrastive learning-guided unified framework with feature selection for anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 9994–10003 (2025)

    2025

  39. [47]

    In: Computer Vision and Pattern Recognition

    Xu,J.,Li,S.,Deng,A.,etal.:Probabilisticknowledgedistillationoffaceensembles. In: Computer Vision and Pattern Recognition. pp. 3489–3498 (2023)

    2023

  40. [48]

    In: European Conference on Computer Vision

    Yao, X., Li, R., Qian, Z., Wang, L., Zhang, C.: Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. In: European Conference on Computer Vision. pp. 92–108. Springer (2024)

    2024

  41. [49]

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

    Yao, X., Li, R., Zhang, J., Sun, J., Zhang, C.: Explicit boundary guided semi- push-pull contrastive learning for supervised anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 24490–24499 (2023)

    2023

  42. [50]

    Advances in Neural Information Processing Systems 35, 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 Systems 35, 4571–4584 (2022)

    2022

  43. [51]

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

    Zavrtanik, V., Kristan, M., Skočaj, D.: Draem-a discriminatively trained re- construction embedding for surface anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 8330–8339 (2021)

    2021

  44. [52]

    arXiv preprint arXiv:2312.07495 (2023)

    Zhang, J., Chen, X., Wang, Y., Wang, C., Liu, Y., Li, X., Yang, M.H., Tao, D.: Exploring plain vit reconstruction for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2312.07495 (2023)

    2023 arXiv

  45. [53]

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

    Zhang, X., Li, S., Li, X., Huang, P., Shan, J., Chen, T.: Destseg: Segmentation guided denoising student-teacher for anomaly detection. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 3914– 3923 (2023)

    2023

  46. [54]

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

    Zhu, J., Ding, C., Tian, Y., Pang, G.: Anomaly heterogeneity learning for open-set supervised anomaly detection. In: Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. pp. 17616–17626 (2024)

    2024

  47. [55]

    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)

    2022

  48. [865]

    Association For Uncertainty in Artificial Intelligence (AUAI) (2018)

    2018

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.