Pith. sign in

REVIEW 3 major objections 4 minor 28 references

LIBAD: A Multimodal Anomaly Detection Benchmark for Li-Ion Battery Electrode Manufacturing

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A new benchmark for Li-ion electrode manufacturing shows defects are often visible in only one modality, and a density-aware memory bank cuts false positives from 60.4% to 54.3%.

desk verdict LIBAD fills a real gap with a useful multimodal battery-manufacturing benchmark, but the benchmark's core validity claim rests on an untested material/batch exchangeability assumption and the headline DA-Core result is selected from a test-set sweep. read the letter →

arxiv 2608.07958 v1 pith:ECZO7IMY submitted 2026-08-08 cs.CV

Wenbo Sui , Daniel Lichau , Harold Phelippeau , Zhao Liu This is my paper · ORCID
classification cs.CV
keywords anomalydetectionmultimodalbenchmarkLi-ionbatteryelectrodesroll-to-rollmanufacturingX-rayradiographyvisible-lightimagingdensity-awarecoresetfalse-positiverate
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

LIBAD is a dataset of 744 electrode patches physically cut from real roll-to-roll Li-ion battery production lines and imaged with double-sided visible light, high-resolution X-ray, and inline-compatible low-resolution X-ray. The paper's central claim is that this is the first multimodal anomaly detection benchmark for continuous process manufacturing, and that it exposes a failure mode missing from object-centric RGB-3D benchmarks: defect evidence is often strong in one modality and weak or absent in another. On the practical inline setting of visible light plus low-resolution X-ray, representative existing methods transfer poorly, with false-positive rates (FPR95, the share of normal samples flagged when 95% of defects are caught) between 59.5% and 69.7%. The paper proposes DA-Core, a memory-based detector whose density-aware coreset selection keeps fine-grained normal variations in a compact memory bank; at a 5% memory ratio it lowers FPR95 from 60.4% to 54.3% and also cuts inference time. If correct, this gives the community a realistic testbed for an underexplored industrial setting and a faster baseline for it.

What carries the argument

The load-bearing mechanism is DA-Core's density-aware farthest point sampling (DAFPS). Standard coreset selection picks the next memory feature farthest from the current bank, which covers the feature space but under-represents dense clusters of common normal textures; DAFPS replaces that score with $q_i = \tilde{r}_i^M (1 + \lambda \tilde{\rho}_i)$, where $\tilde{r}_i^M$ is the normalized distance to the current memory bank and $\tilde{\rho}_i$ is a log-transformed, quantile-normalized local density estimate from a Gaussian kernel over the $k$ nearest neighbors. This keeps global coverage while favoring representative variations inside dense normal regions, so normal test patches are more likely to find close memory neighbors. The benchmark itself is also a named central object: LIBAD, a set of 744 physically cropped electrode patches with aligned visible-light, high-resolution X-ray, and low-resolution X-ray images, registered by optimizing a correlation-ratio similarity measure and organized into 10 stratified splits.

What would settle it

Train DA-Core separately on anode and cathode material groups (or on one production batch) and test on the other; if the FPR95 on normal patches from the unseen group jumps far above the within-group value while defect detection holds, then the memory bank is encoding material- or batch-specific texture rather than defect-related evidence, and the benchmark's anomaly scores partly measure domain shift.

Watch

Extended reading notes

Core claim

The paper establishes two related results. First, electrode patches in LIBAD are visually homogeneous, and among the 383 anomalous samples 29.5% are visible in only one modality (11.0% visible only in visible light and 18.5% only in X-ray), so anomaly evidence is modality-selective rather than shared. Second, under the practical inline setting of visible light plus low-resolution X-ray, existing multimodal methods that assume strong cross-modal correspondence or RGB-3D priors transfer poorly, with FPR95 across baselines ranging from 59.5% to 69.7%. DA-Core, which selects memory features by balancing global coverage with local density of normal features, reaches 54.3% FPR95 at a 5% coreset ratio, outperforming the best standard farthest-point-sampling result obtained at a 20% ratio while reducing inference time by 43.9%. The authors conclude that both the distribution of normal features and the relationship between modalities need explicit treatment in process-manufacturing anomaly detection.

Load-bearing premise

The benchmark's validity rests on the assumption that normal patches cut near defective regions or under the same production conditions differ from anomalous patches only by the defect, so that after pooling, anomaly scores measure defect evidence rather than material, batch, or process-setting differences.

Editorial extensions

If this is right

  • If LIBAD is a fair testbed, then multimodal anomaly detection for continuous processes cannot be evaluated on object-centric RGB-3D benchmarks alone.
  • If DA-Core's improvement is real, density-aware coreset selection should become the default starting point for memory-based industrial anomaly detection with non-uniform normal feature distributions.
  • At the reported 5% memory ratio, DA-Core's FPR95 (54.3%) beats the best standard FPS result (56.6% at 20% ratio) while cutting inference time by 43.9%, so compact memory banks need not sacrifice false-positive control.
  • Because the absolute FPR95 remains around 54%, the paper's own conclusion is that LIBAD is still an open benchmark and no tested method is deployment-ready.

Reading between the lines

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

  • Beyond the paper's claims, an unresolved question is how much of the measured anomaly signal is defect evidence versus material- or batch-specific texture; a held-out-batch evaluation, training on one production batch or material and testing on another, would isolate this.
  • The modality-selective subset (42 visible-only, 71 X-ray-only anomalies) suggests a direct test for fusion rules: a method that knows when to down-weight a modality should beat one that always trusts both, and per-defect-category scores would show where false positives concentrate.
  • Density-aware selection is a generic trick that could plausibly transfer to any memory-based anomaly detector on homogeneous textured surfaces, including other continuous processes such as foil, film, or textile manufacturing.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper presents LIBAD, a multimodal anomaly detection benchmark for Li-ion battery electrode manufacturing, containing 744 registered electrode patches (VIS, high-resolution X-ray, inline low-resolution X-ray) with 11 defect categories and image-level labels. The authors benchmark several existing multimodal methods under a VIS+X-rayL protocol and propose DA-Core, a PatchCore-style memory-bank method whose coreset selection adds a local-density term to farthest point sampling. They report that at a 0.05 coreset ratio DA-Core lowers FPR95 from 60.4% to 54.3% relative to standard FPS, and that it reaches lower FPR95 than the best standard FPS result while reducing inference time. They also quantify cross-modal anomaly inconsistency, with 29.5% of anomalous samples visible in only one of the two analyzed modalities.

Significance. If the dataset is valid, LIBAD is a useful contribution: it extends industrial anomaly detection to continuous process manufacturing and to weakly correlated modalities, using real production data with aligned imaging and an inline-compatible evaluation protocol. The paper is careful in several ways: 10 official splits, registration-quality verification, full acquisition details, and an honest acknowledgment that absolute FPR95 remains too high for deployment. DA-Core is a modest but plausible improvement in false-positive control, and the density-aware coreset idea is simple and reproducible. The main caveat is that the benchmark's validity, and therefore the meaningfulness of the method comparison, depends on controlling material and batch confounds that are not yet demonstrated.

major comments (3)
  1. [Dataset Construction / Appendix Note 1 / Table S1] The central validity claim of the benchmark—that anomaly scores reflect defect-related evidence rather than material, batch, or process differences—rests on an exchangeability assumption that is stated but not verified. The only safeguard offered is that normal patches were collected 'near the defective samples or under the same manufacturing conditions whenever possible,' yet Table S1 shows the 744 samples span anode, NMC/NCM, LFP, LCO, and unspecified cathode categories, and Note 1 states that exact formulations are commercially confidential. No per-sample material or batch metadata is released, no cross-tabulation of material versus defect group is provided, and no analysis demonstrates that normal and anomalous test patches within each group are matched in composition and process conditions. Because all normal samples are pooled during training, a defect category that is disproportionately associated with one material or one production run could be detected largely by global material appearance rather than by localized defect evidence, inflating AUROC and deflating FPR95. This is load-bearing for both the benchmark and the DA-Core comparison. Please add group-level or material-matched results (e.g., per-group AUROC/FPR95, a matched-pair analysis, or anonymized batch/material metadata) to support the exchangeability claim.
  2. [Experiments (Ablation, Table 3)] The headline improvement (FPR95 60.4% to 54.3% at coreset ratio 0.05) is obtained with lambda = 0.7, selected as the best value from a sensitivity scan evaluated on the same 10 test splits; no validation-based selection is described. Choosing the density weight from the test data is a form of test-set fitting and makes the reported gain optimistically biased. Although the improvement is not isolated to lambda = 0.7 (lambda = 0.3 already gives 57.1%), the central quantitative claim should either use a pre-specified or nested selection rule, or report a scalar summary over the sensitivity table. Please clarify how lambda (and k, alpha, beta) were set and, if possible, use the official validation splits for model selection.
  3. [Experiments (Table 2, Eq. (10))] The comparison between DA-Core and the PatchCore VIS+X-rayL baseline is not fully controlled, because the manuscript does not state how the multimodal PatchCore baseline fuses VIS and X-rayL, while DA-Core adds an OCSVM late-fusion module with scaling factors alpha and beta. If PatchCore uses simple score averaging or feature concatenation, part of DA-Core's improvement over PatchCore in Table 2 could come from the fusion module rather than from density-aware coreset selection. The ablation in Fig. 3(a,b) appears to compare standard FPS with density-aware FPS while keeping the rest of the pipeline fixed, which is the correct controlled comparison; please state explicitly that this is the case and specify the PatchCore fusion procedure.
minor comments (4)
  1. [Method (Eq. (6))] The value of k and the choice of tau are not specified; report them in the implementation details.
  2. [Method (Eq. (10))] The values of alpha and beta and the OCSVM hyperparameters are not specified; they should be listed to make the fusion module reproducible.
  3. [Experiments (Table 2)] The paper reports standard deviations but no paired significance tests or confidence intervals for the FPR95 differences; given the overlap in standard deviations (e.g., 54.3 +/- 5.7 vs. 60.4 +/- 3.9), a paired test would strengthen the claim that the improvement is not noise.
  4. [Experiments (Fig. 3)] The caption of Fig. 3 should state explicitly whether the 'standard FPS' curve uses the same OCSVM fusion and the same anomaly-scoring pipeline as DA-Core, so that the controlled comparison is unambiguous.

Circularity Check

1 steps flagged · score 4.0 of 10

No structural circularity in the benchmark construction; one headline quantitative claim (DA-Core's λ=0.7 FPR95) is selected on the test splits, giving a mildly fitted result.

  1. fitted input called prediction [Experiments, Ablation Study — 'Sensitivity to density weight' (Table 3) and final DA-Core row in Table 2]
    "Table 3 reports the mean and standard deviation of FPR95 over the 10 official splits. Without density weighting, the model obtains 60.4% FPR95. Introducing density awareness consistently reduces FPR95, with the best result achieved at λ=0.7 (54.3%)."

    The paper presents DA-Core's headline FPR95 of 54.3% after sweeping the density weight λ and taking the best value (λ=0.7) from Table 3. No validation-based selection of λ is described; the 54.3% is the minimum of the λ grid evaluated on the same 10 test splits used for all reported results. Thus the central 'prediction' that DA-Core reduces FPR95 from 60.4% to 54.3% is partly a fitted optimum of a hyperparameter chosen on test labels, not an independent fixed-model evaluation. The qualitative direction holds at every tested λ>0, so this is partial selection bias rather than a derivation that equals its input; it does not make the coreset objective itself circular.

full rationale

LIBAD's dataset construction and the DA-Core coreset objective (Eq. 8) are not derived from the benchmark outcomes: the density-aware selection rule is a defined algorithm and its comparison with standard FPS is an empirical measurement. The OCSVM late-fusion component cites prior work by Sui et al. (CMDIAD), but it is presented alongside M3DM as a design choice and is not load-bearing for the benchmark claim. The main validity concern, namely the exchangeability of pooled normal patches across anode and cathode material categories and production conditions (Appendix Table S1, Note 1), is a real confounding/correctness risk but not circularity, because defect status is not defined in terms of material membership. The only concrete circularity-adjacent step is the test-set selection of λ, which partly optimizes one headline number but does not invalidate the method's qualitative improvement or the dataset's independent content.

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

The paper relies on standard unsupervised anomaly detection assumptions, plus the adequacy of its DINOv3 features and manual labels. The only data-fitted quantities are hyperparameters (lambda, k, alpha, beta), some unreported. No new physical entities are introduced.

free parameters (3)
  • lambda (density weight) = 0.7
    Eq. 8 balances distance coverage and local density. Value chosen from a test-set sensitivity sweep (Table 3); no validation-based selection is stated.
  • k (nearest-neighbor count for density) = not reported
    Eq. 6 defines density using k nearest neighbors, but the paper never states k's value or selection rule.
  • alpha, beta (modality scaling factors) = not reported
    Eq. 10 uses modality scaling factors in OCSVM fusion; no values or fitting procedure are given.
assumptions (3)
  • domain assumption Normal training patches are representative of the normal population and are free of defect-related confounds from material, batch, or process differences.
    Pooling normals from different defect groups assumes exchangeability. Acknowledged in Dataset Construction but not quantitatively validated.
  • domain assumption Frozen DINOv3 features capture the normal variations relevant to electrode defects.
    Feature extraction uses DINOv3 ViT-S/16; the method's scoring inherits whatever representational biases that backbone has.
  • domain assumption Manual defect labels and modality-visibility assessment are accurate enough to serve as ground truth.
    Labels come from process engineers and joint experts, with re-annotation, but no inter-annotator agreement or external validation is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LIBAD: A Multimodal Anomaly Detection Benchmark for Li-Ion Battery Electrode Manufacturing." pith.science (2026). https://pith.science/paper/ECZO7IMY

@misc{pith2026260807958,
  author       = {Pith},
  title        = {Pith review of: LIBAD: A Multimodal Anomaly Detection Benchmark for Li-Ion Battery Electrode Manufacturing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ECZO7IMY}},
  note         = {Machine review of arXiv:2608.07958}
}
read the original abstract

Multimodal industrial anomaly detection has largely focused on discrete products using strongly correlated RGB and 3D observations, leaving continuous process manufacturing and weakly correlated sensing modalities underexplored. We introduce LIBAD, the first multimodal anomaly detection benchmark for Li-ion battery electrode manufacturing. Collected from real roll-to-roll production lines, LIBAD provides aligned double-sided visible-light imaging, high-resolution X-ray radiography, and inline-compatible low-resolution X-ray radiography. Electrode patches in LIBAD exhibit highly homogeneous material appearance, while defect evidence can be strong in one modality but weak or absent in another, resulting in pronounced cross-modal anomaly inconsistency. Benchmarks of representative methods under the inline-compatible visible-light and low-resolution X-ray setting exhibit limited transferability and consistently high false-positive rates. We therefore propose DA-Core, a memory-based method that jointly considers feature-space coverage and local density of normal features during coreset selection, allowing compact memory banks to better preserve fine-grained normal variations. With a coreset ratio of 0.05, DA-Core reduces FPR95 from 60.4% to 54.3% compared with standard farthest point sampling. At this ratio, DA-Core also outperforms the best standard coreset result (obtained at 0.20) while reducing inference time by 43.9%. These results suggest that both the data distribution of normal features and the modality relationship itself require explicit consideration when designing anomaly detection methods for process manufacturing.

Figures

Figures reproduced from arXiv: 2608.07958 by the authors.

Figure 1
Figure 1. Example defects in LIBAD across four visibility [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of DA-Core. For X-rayL, patch tokens are extracted by a separate encoder and form the X-rayL memory candidate set: Z x = [ i Fx(x x i ). (3) Density-aware Coreset Construction For each modality m ∈ {v, x}, let Z m denote the full can￾didate set. The goal of coreset selection is to construct a compact memory bank Mm ⊂ Zm while preserving repre￾sentative normal patterns. Standard farthest point sampling selec… view at source ↗
Figure 3
Figure 3. Quantitative and qualitative comparison of standard FPS and density-aware FPS. (a) FPR95 and (b) AUROC under [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 25 canonical work pages

  1. [1]

    M.; Moch, E.; and Herring, P

    Attia, P. M.; Moch, E.; and Herring, P. K. 2025. Chal- lenges and opportunities for high-quality battery pro- duction at scale.Nature Communications, 16: 611

    work page 2025

  2. [2]

    2019.MVTecAD–Acomprehensivereal-worlddataset for unsupervised anomaly detection

    Bergmann,P.;Fauser,M.;Sattlegger,D.;andSteger,C. 2019.MVTecAD–Acomprehensivereal-worlddataset for unsupervised anomaly detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9592–9600

    work page 2019

  3. [3]

    Bergmann, P.; Jin, X.; Sattlegger, D.; and Steger, C

    work page

  4. [4]

    Bonfiglioli, L.; Toschi, M.; Silvestri, D.; Fioraio, N.; and De Gregorio, D. 2022. The eyecandies dataset for unsupervised multimodal anomaly detection and localization. InProceedings of the Asian Conference on Computer Vision, 3586–3602

    work page 2022

  5. [5]

    Cohen, N.; and Hoshen, Y. 2020. Sub-image anomaly detection with deep pyramid correspondences.arXiv preprint arXiv:2005.02357

    arXiv 2020

  6. [6]

    InProceedings of the IEEE/CVFConferenceonComputerVisionandPattern Recognition, 17234–17243

    Costanzino,A.;Ramirez,P.Z.;Lisanti,G.;andDiSte- fano,L.2024.Multimodalindustrialanomalydetection by crossmodal feature mapping. InProceedings of the IEEE/CVFConferenceonComputerVisionandPattern Recognition, 17234–17243

    work page 2024

  7. [7]

    Defard, T.; Setkov, A.; Loesch, A.; and Audigier, R

    work page

  8. [8]

    Guo, Y.; Zhang, W.; and Wang, X. 2026. Remove the Ambiguity: Few-shot Multimodal Anomaly Detection Using Crossmodal Feature Replacers. InForty-third International Conference on Machine Learning

    work page 2026

Show all 28 references
  1. [9]

    Springer

    Padim:apatchdistributionmodelingframework foranomalydetectionandlocalization.InInternational conference on pattern recognition, 475–489. Springer

  2. [10]

    HGCF: Hierarchical Geometry-Color Fusion for Mul- timodalIndustrialAnomalyDetection

    Li,M.;He,J.;Li,J.;Han,D.;Wan,J.;andLi,G.2025. HGCF: Hierarchical Geometry-Color Fusion for Mul- timodalIndustrialAnomalyDetection. InProceedings ofthe33rdACMInternationalConferenceonMultime- dia, 1023–1031

    2025

  3. [11]

    A.; Widmer, J

    Lambert,S.M.;Armstrong,M.;Attidekou,P.S.;Chris- tensen, P. A.; Widmer, J. D.; Wang, C.; and Scott, K

  4. [12]

    Li, Z.; Yan, Y.; Wang, X.; Ge, Y.; and Meng, L. 2025. Asurveyofdeeplearningforindustrialvisualanomaly detection.Artificial Intelligence Review, 58(9): 279

    2025

  5. [13]

    Lin,Y.;Chang,Y.;Tong,X.;Yu,J.;Liotta,A.;Huang, G.; Song, W.; Zeng, D.; Wu, Z.; Wang, Y.; and Zhang, W. 2025. A survey on RGB, 3D, and multimodal ap- proaches for unsupervised industrial image anomaly detection.Information Fusion, 121: 103139

    2025

  6. [14]

    Li, W.; Zheng, B.; Xu, X.; Gan, J.; Lu, F.; Li, X.; Ni, N.; Tian, Z.; Huang, X.; Gao, S.; et al. 2025. Multi- sensorobjectanomalydetection:Unifyingappearance, geometry,andinternalproperties.InProceedingsofthe computer vision and pattern recognition conference, 9984–9993

    2025

  7. [15]

    Luo, A.; Wen, G.; Cheng, Y.; Mei, S.; Dong, H.; and Liu, X. 2024. DMMGNet: A discrimination mapping and memory bank mean guidance-based network for high-performance few-shot industrial anomaly detec- tion.Neurocomputing, 610: 128622

    2024

  8. [16]

    Ma, H.; Yang, G.; Zhao, D.; Ji, Y.; and Zuo, W. 2025. ReMP-AD: Retrieval-enhanced Multi-modal Prompt Fusion for Few-Shot Industrial Visual Anomaly De- tection. InProceedingsoftheIEEE/CVFInternational Conference on Computer Vision, 20425–20434

    2025

  9. [17]

    Liu, J.; Xie, G.; Wang, J.; Li, S.; Wang, C.; Zheng, F.; and Jin, Y. 2024. Deep industrial image anomaly detection: A survey.Machine Intelligence Research, 21(1): 104–135

    2024

  10. [18]

    InProceedings of the IEEE/CVF conferenceoncomputervisionandpatternrecognition, 14318–14328

    Roth, K.; Pemula, L.; Zepeda, J.; Schölkopf, B.; Brox, T.;andGehler,P.2022.Towardstotalrecallinindustrial anomaly detection. InProceedings of the IEEE/CVF conferenceoncomputervisionandpatternrecognition, 14318–14328

    2022

  11. [19]

    V.; Seitzer, M.; Baldassarre, F.; Oquab, M.; Jose, C.; Khalidov, V.; Szafraniec, M.; Yi, S.; Ramamonjisoa, M.; et al

    Siméoni, O.; Vo, H. V.; Seitzer, M.; Baldassarre, F.; Oquab, M.; Jose, C.; Khalidov, V.; Szafraniec, M.; Yi, S.; Ramamonjisoa, M.; et al. 2025. Dinov3.arXiv preprint arXiv:2508.10104

    2025 arXiv

  12. [20]

    Roche, A.; Malandain, G.; Pennec, X.; and Ayache, N. 1998. The correlation ratio as a new similarity measure for multimodal image registration. In Wells, W. M.; Colchester, A.; and Delp, S., eds.,Medical Im- ageComputingandComputer-AssistedIntervention— MICCAI’98, 1115–1124. S...

    1998

  13. [21]

    Tao, C.; Cao, X.; and Du, J. 2025. G2SF: Geometry-Guided Score Fusion for Multimodal In- dustrial Anomaly Detection. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, 20551–20560

    2025

  14. [22]

    Wang, Y.; Peng, J.; Zhang, J.; Yi, R.; Wang, Y.; and Wang, C. 2023. Multimodal industrial anomaly detec- tionviahybridfusion. InProceedingsoftheIEEE/CVF conferenceoncomputervisionandpatternrecognition, 8032–8041

    2023

  15. [23]

    Sui, W.; Lichau, D.; Lefèvre, J.; and Phelippeau, H

  16. [24]

    Real-iadd3:Areal-world2d/pseudo-3d/3d dataset for industrial anomaly detection

    Zhu, W.; Wang, L.; Zhou, Z.; Wang, C.; Pan, Y.; Zhang, R.; Chen, Z.; Cheng, L.; Gao, B.-B.; Zhang, J.; etal.2025. Real-iadd3:Areal-world2d/pseudo-3d/3d dataset for industrial anomaly detection. InProceed- ings of the Computer Vision and Pattern Recognition Conference, 15214–15...

    2025

  17. [27]

    Zhao, Y.; Zhang, X.; Lu, H.; and Zhang, L. 2026. Complementary Prototype Mapping for Efficient Mul- timodal Anomaly Detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, 14178–14187

    2026

  18. [2017]

    Rapid Nondestructive-Testing Technique for In- Line Quality Control of Li-Ion Batteries.IEEE Trans- actions on Industrial Electronics, 64(5): 4017–4026

  19. [2021]

    The mvtec 3d-ad dataset for unsupervised 3d anomaly detection and localization.arXiv preprint arXiv:2112.09045

    arXiv

  20. [2025]

    Incomplete multimodal industrial anomaly de- tection via cross-modal distillation.Information Fu- sion, 103572

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.