The DOI matches its Crossref record; 50 of the 55 references it cites were found in a registry or an index; none of the 4 sources its claims cite is checked for support yet.
The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection
Open accessAbstract
The detection of anomalous structures in natural image data is of utmost importance for numerous tasks in the field of computer vision. We introduce the MVTec anomaly detection dataset containing 5354 high-resolution color images of different object and texture categories. It contains normal, i.e., defect-free images intended for training and images with anomalies intended for testing. The anomalies manifest themselves in the form of over 70 different types of defects such as scratches, dents, contaminations, and various structural changes. In addition, we provide pixel-precise ground truth annotations for all anomalies. This benchmark indicates that methods that leverage descriptors of pretrained networks outperform all other approaches and deep-learning-based generative models show considerable room for improvement.
Overview
CoChat Summary
TL;DR
The paper introduces MVTec AD, a 5354-image dataset for unsupervised anomaly detection with pixel-precise labels, and benchmarks seven methods on it. It finds that methods using pretrained-network features, especially Student-Teacher, outperform from-scratch generative models, while threshold selection remains highly challenging.
Key findings
- MVTec AD contains 5354 high-resolution images across 15 categories: 3629 defect-free training images, 1725 test images (467 good, 1258 defective), 73 defect types, and 1888 pixel-precise anomalous regions.
- Student-Teacher anomaly detection performs best overall, with mean AU-PR 0.599, AU-ROC 0.922, AU-PRO 0.924, and AU-IoU 0.190; CNN Feature Dictionary is second (AU-PR 0.466, AU-ROC 0.836, AU-PRO 0.803, AU-IoU 0.168).
- f-AnoGAN performs worst across the threshold-independent metrics (mean AU-PR 0.136, AU-ROC 0.472, AU-PRO 0.528, AU-IoU 0.073), and deep generative models trained from scratch show considerable room for improvement.
- Only 2.7% of test-set pixels are anomalous; the paper reports that threshold selection on anomaly-free validation data is highly challenging and that metric rankings can change with the FPR integration limit.
Methods
- The authors introduce the MVTec AD dataset, covering five texture categories and ten object categories, with images acquired using a 2048 × 2048 industrial RGB sensor and cropped to 700 × 700–1024 × 1024 pixels; three categories are grayscale.
- They benchmark f-AnoGAN, CNN Feature Dictionary, Student-Teacher anomaly detection, ℓ2-autoencoder, SSIM-autoencoder, GMM-based Texture Inspection, and Variation Model.
- Evaluation uses pixel-level and region-level metrics including AU-PR, AU-ROC, AU-PRO, and AU-IoU, plus threshold estimation techniques (Maximum, p-Quantile, k-Sigma, Max-Area) estimated solely on anomaly-free validation images.
Limitations
- The paper states that selecting a suitable threshold purely on anomaly-free validation images is highly challenging, and that the same estimator can yield very different results depending on method and dataset category.
- The paper states that runtime and memory consumption are not reported as exact numbers because they depend heavily on implementation details, libraries, and parallelization.
- The paper states that no evaluated method excels across the entire dataset, and that methods perform unevenly across object and defect categories; the text also makes evident that the dataset is limited to 15 industrial inspection categories with manually generated defects and no multiple poses per physical sample.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
0 Valid · 5 Not valid · 2 Not checked · 114 Own results and interpretations
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“We introduce the MVTec anomaly detection dataset containing 5354 high-resolution color images of different object and texture categories.”
- Also stated in: § 1 Introduction · p. 2 · plain
“consists of 5354 high-resolution images of five unique textures and ten unique objects from different domains.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § 1 Introduction · p. 2 · plain
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“However, in some cases, it entirely fails to produce any true positives as well, due to outliers on the validation set.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Methods that require only a single model evaluation per image and run entirely on the GPU, such as the autoencoders evaluated on the objects of the dataset, allow for much faster inference times in the range of a few milliseconds.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Our results show that discriminative approaches that leverage descriptors of pretrained networks outperform methods that learn feature representations from scratch solely on the anomaly-free training data.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Furthermore, the marker locations of the two thresholds are often very different for the same anomaly detection method, which indicates that the assumption of normally distributed anomaly scores often does not hold in practice.”
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
64 more claims
Create a free account to see all 121 claims and every source verdict.
- It may state two separate claims as one. 32
- Probably not a substantial claim. 23
- The quote only partly states this claim. 3
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“The vast majority (97%) of anomalies in the dataset matches our definition of structural anomalies.Hence, an evaluation on this dataset alone does not give sufficient insight into how well a method detects logical anomalies.”
From Beyond Dents and Scratches: Logical Constraints in Unsupervised Anomaly Detection and Localization (2022) · cites this paper as Bergmann et al. 2021
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“We find that existing datasets are biased towards local structural anomalies such as scratches, dents, or contaminations.”
From Beyond Dents and Scratches: Logical Constraints in Unsupervised Anomaly Detection and Localization (2022) · cites this paper as Bergmann et al. 2021
Experimental — accuracy not yet measuredThe check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“In particular, they lack anomalies in the form of violations of logical constraints, e.g., permissible objects occurring in invalid locations.”
From Beyond Dents and Scratches: Logical Constraints in Unsupervised Anomaly Detection and Localization (2022) · cites this paper as Bergmann et al. 2021
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ AAdditional Results · p. 17
“Tables 5, 6 and 7 report the area under the ROC, IoU, and PR curves, respectively.”
§ AAdditional Results · p. 18
“Table 6 Normalized area under the PRO curve up to an average false positive rate per-pixel of 30% for each dataset category”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 5 Threshold Selection · p. 9
“In this work, we consider three threshold estimation techniques for anomaly segmentation where the thresholds are estimated solely on a set of anomaly-free validation images prior to testing.”
§ 5 Threshold Selection · p. 10
“Max-Area Threshold All estimators presented so far compute thresholds simply on the one-dimensional distribution of validation anomaly scores”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Abstract · p. 1
“We introduce the MVTec anomaly detection dataset containing 5354 high-resolution color images of different object and texture categories.”
§ 3 Description of the Dataset · p. 7
“Since gray-scale images are also common in industrial inspection, three object categories ( grid , screw , and zipper ) are made available solely as single-channel images.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 3 Description of the Dataset · p. 7
“three object categories ( grid , screw , and zipper )”
§ 2.2.1 Generative Adversarial Networks (GANs) · p. 5
“| Textures | Grid | 264 | 21 | 57 | 5 | 170 | 1024 | ✓ |”
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474 · cited 4×
“Napoletano et al. (2018) propose to use clustered feature descriptions obtained from the activations of a ResNet-18 (He et al. 2016) classification network pretrained on Ima-geNet to distinguish normal from anomalous data.”
MethodInfluentialImageNet classification with deep convolutional neural networks
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017) · Communications of the ACM · doi:10.1145/3065386 · Cited by 112,155 · cited 4×
“1998), CIFAR10 (Krizhevsky and Hinton 2009), and ImageNet (Krizhevsky et al. 2012).”
BackgroundMethodInfluentialGradient-based learning applied to document recognition
Lecun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998) · Proceedings of the IEEE · doi:10.1109/5.726791 · Cited by 58,805
Image quality assessment: from error visibility to structural similarity
Wang, Z., Bovik, A. C., Sheikh, H. R., & Simoncelli, E. P. (2004) · IEEE Transactions on Image Processing · doi:10.1109/tip.2003.819861 · Cited by 57,892
Learning Multiple Layers of Features from Tiny Images
Krizhevsky, A. (2024) · doi:10.57702/zp44cu3g · Cited by 23,395
Auto-Encoding Variational Bayes
Kingma, D. P., & Welling, M. (2013) · arXiv · doi:10.48550/arxiv.1312.6114 · Cited by 20,067 · cited 2×
“There exist various extensions to CAEs such as memory-augmented (Gong et al. 2019) or variational autoencoders (VAEs) (Kingma and Welling 2014).”
BackgroundGAN(Generative Adversarial Nets)
(2017) · Journal of Japan Society for Fuzzy Theory and Intelligent Informatics · doi:10.3156/jsoft.29.5_177_2 · Cited by 19,270
David, F. N., & Tukey, J. W. (1977) · Biometrics · doi:10.2307/2529486 · Cited by 12,708
The Cityscapes Dataset for Semantic Urban Scene Understanding
Cordts, M., Omran, M., Ramos, S., Rehfeld, T., Enzweiler, M., Benenson, R., Franke, U., Roth, S., & Schiele, B. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.350 · Cited by 12,433 · cited 2×
“They artificially inserted images of novel objects into images of the Cityscapes dataset (Cordts et al. 2016), for which pixel-precise annotations are available.”
Method(2016) · Cited by 11,679
Research works printed but not linked: 2
- Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014). Generative adversarial nets. In Advances in neural information processing systems (pp. 2672–2680).
- Steger, C. (2002). Occlusion, clutter, and illumination invariant object recognition. International Archives of Photogrammetry and Remote Sensing, vol XXXIV, part, 3A, 345–350.
Cited by
SPot-the-Difference Self-supervised Pre-training for Anomaly Detection and Segmentation
Zou, Y., Jeong, J., Pemula, L., Zhang, D., & Dabeer, O. (2022) · Lecture Notes in Computer Science · doi:10.1007/978-3-031-20056-4_23 · Cited by 651
Deep Industrial Image Anomaly Detection: A Survey
Liu, J., Xie, G., Wang, J., Li, S., Wang, C., Zheng, F., & Jin, Y. (2024) · Machine Intelligence Research · doi:10.1007/s11633-023-1459-z · Cited by 463
“Per-Region Overlap (PRO) [154] ↑ PRO = 1 N ∑”
EfficientAD: Accurate Visual Anomaly Detection at Millisecond-Level Latencies
Batzner, K., Heckler, L., & König, R. (2024) · 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) · doi:10.1109/wacv57701.2024.00020 · Cited by 386
“[7] evaluate some of these methods and find that they fail when requirements such as wellaligned objects are not met.”
BackgroundMethodResultInfluentialSelf-citationBeyond Dents and Scratches: Logical Constraints in Unsupervised Anomaly Detection and Localization
Bergmann, P., Batzner, K., Fauser, M., Sattlegger, D., & Steger, C. (2022) · International Journal of Computer Vision · doi:10.1007/s11263-022-01578-9 · Cited by 286
“However, as motivated in Bergmann et al. (2021), we discourage an evaluation of our dataset at high false positive rates since those correspond to segmentation results that are no longer meaningful in practice.”
BackgroundMethodInfluentialSelf-citationComputer Vision Techniques in Manufacturing
Zhou, L., Zhang, L., & Konz, N. (2023) · IEEE Transactions on Systems, Man, and Cybernetics: Systems · doi:10.1109/tsmc.2022.3166397 · Cited by 248
“MVTec is a recent commonly used dataset for benchmarking anomaly detection methods with a focus on industrial inspection [155], [156].”
Background
Metadata
- Authors
- Bergmann, P., Batzner, K., Fauser, M., Sattlegger, D., & Steger, C.
- Affiliations
- Technical University of Munich (DE)
- Corresponding author
- Bergmann, Paul
- Venue
- International Journal of Computer Vision
- Published
- 2021-01-06
- Volume
- 129
- Issue
- 4
- Pages
- 1038-1059
- ISSN
- 0920-5691 (ISSN-L)
- 1573-1405
- OpenAlex
- W3118600296
- Semantic Scholar
- 48f9a48aa5b1230b05a443d2d531e6441a541686
- Cited by
- 602
- References
- 55
- Citations by year
- 2026: 187 · 2025: 148 · 2024: 113 · 2023: 83 · 2022: 54 · 2021: 17
- FWCI
- 34.45
- Type
- journal article
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- creativecommons.org/licenses/by/4.0 (tdm)
- creativecommons.org/licenses/by/4.0 (vor)
- Locations
- International Journal of Computer Vision (publishedVersion, cc-by)
- mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) (submittedVersion, other-oa)
- Topics
- Physical Sciences › Engineering › Industrial and Manufacturing Engineering › Industrial Vision Systems and Defect Detection
- Physical Sciences › Computer Science › Artificial Intelligence › Anomaly Detection Techniques and Applications
- Physical Sciences › Engineering › Civil and Structural Engineering › Infrastructure Maintenance and Monitoring
- Keywords
- unsupervised anomaly detection, image defects, ground truth annotation, convolutional autoencoder, generative adversarial networks, pre-trained convolutional neural networks, feature descriptors, benchmark evaluation, threshold estimation, performance indicators
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection. https://app.cochat.ai/papers/10.1007/s11263-020-01400-4
- The paper
- Bergmann, P., Batzner, K., Fauser, M., Sattlegger, D., & Steger, C. (2021). The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection. International Journal of Computer Vision. https://doi.org/10.1007/s11263-020-01400-4
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.