The DOI matches its Crossref record; 51 of the 51 references it cites were found in a registry or an index; 13 of the 27 sources its claims cite were checked for support.
A Unified Model for Multi-Class Anomaly Detection
Open accessAbstract
Despite the rapid advance of unsupervised anomaly detection, existing methods require to train separate models for different objects. In this work, we present UniAD that accomplishes anomaly detection for multiple classes with a unified framework. Under such a challenging setting, popular reconstruction networks may fall into an 'identical shortcut', where both normal and anomalous samples can be well recovered, and hence fail to spot outliers. To tackle this obstacle, we make three improvements, including a layer-wise query decoder to help model the multi-class distribution. When learning a unified model for 15 categories in MVTec-AD, we surpass the second competitor on the tasks of both anomaly detection (from 88.1% to 96.5%) and anomaly localization (from 89.5% to 96.8%).
Overview
CoChat Summary
TL;DR
UniAD is a unified feature-reconstruction framework for multi-class anomaly detection that avoids the “identical shortcut” via a layer-wise query decoder, neighbor masked attention, and feature jittering. On MVTec-AD’s 15-category unified setting it reaches 96.5% AUROC for anomaly detection and 96.8% for localization, substantially outperforming prior methods.
Key findings
- On MVTec-AD with one unified model for 15 categories, UniAD achieves 96.5% AUROC for anomaly detection versus 88.1% for the second competitor DRAEM, and 96.8% for anomaly localization versus 89.5% for the second competitor PaDiM.
- On CIFAR-10 under the proposed many-versus-many unified setting, UniAD reaches 87.2% mean AUROC, outperforming FCDD+OE by 8.3% without auxiliary outlier exposure; baselines US and FCDD fall to 55.9% and 55.8% mean AUROC.
- Ablations show the query embedding is critical: without query embedding, detection/localization AUROC is 69.5%/79.4%; with one query embedding it rises to 87.6%/92.8%; with layer-wise query it reaches 95.0%/96.5%; combining layer-wise query, neighbor masked attention, and feature jittering gives 96.5%/96.8%.
- Transformer-based competitors under the unified setting perform far worse: InTra 65.3%/70.6%, VT-ADL 55.4%/64.4%, AnoViT 69.6%/68.4%, while a UniAD baseline with one query embedding already reaches 87.6%/92.8%.
Methods
- UniAD reconstructs features from a fixed ImageNet-pretrained EfficientNet-b4 backbone: stage-1 to stage-4 features are resized, concatenated into a 272-channel map, tokenized, and linearly projected to 256 channels; the model uses a Neighbor Masked Encoder and Layer-wise Query Decoder, with neighbor masked attention masking each token’s self and neighbors, and feature jittering adding Gaussian noise to input tokens.
- Training on MVTec-AD uses 224×224 images, 14×14 feature size, 1000 epochs on 8 NVIDIA Tesla V100 16GB GPUs, batch size 64, AdamW with weight decay 1×10⁻⁴, initial learning rate 1×10⁻⁴ dropped by 0.1 after 800 epochs, 4 encoder and 4 decoder layers, neighbor size 7×7, jittering scale 20, jittering probability 1, and 5 random seeds; CIFAR-10 uses stage-1 to stage-5 features forming a 720-channel map and batch size 128.
- The objective is MSE feature reconstruction loss; anomaly localization scores are the L2 norm of reconstruction differences, upsampled to image size, and image-level anomaly detection takes the maximum of the average-pooled score map; evaluation uses AUROC under the unified case without fine-tuning.
Limitations
- Stated in the paper: the method does not use category labels, and incorporating labels into the unified model should be further studied; practical normal samples are not as consistent as MVTec-AD, though UniAD handles all 15 MVTec-AD categories; anomaly detection for video surveillance may infringe personal privacy.
- Evident from the text: evaluation is limited to MVTec-AD and CIFAR-10, with no real-world video-surveillance experiments despite the privacy discussion; the unified setting assumes class labels are inaccessible at both training and inference, so the method cannot exploit category information that might improve multi-class modeling.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
8 Valid · 5 Not valid · 6 Not checked · 95 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Despite the rapid advance of unsupervised anomaly detection, existing methods require to train separate models for different objects.”
- Also stated in: § 1 Introduction · p. 1 · plain
“existing methods [6, 11, 26, 28, 49, 50, 53] propose to train separate models for different classes of objects”
What the sources say
- Not checked yetcites [6] [6] P. Bergmann, M. Fauser, D. Sattlegger, and C. Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In IEEE Conf. Comput. Vis. Pattern Recog. , 2020. - source
Checked before — its earlier verdict is held back until a new check runs.
- Partly supportedcites [11] [11] T. Defard, A. Setkov, A. Loesch, and R. Audigier. PaDim: A patch distribution modeling framework for anomaly detection and localization. In Int. Conf. Pattern Recog. , 2021. - sourceExperimental — accuracy not yet measured
The Experimental setups section (p. 4) says the paper's VAE baseline is trained on each MVTec AD class, supporting class-specific training as an example. But the Scalability gain section (p. 6) says SPADE requires no training, so the paper does not support the broader claim that existing methods generally require separate models for different objects. (LLM, conf=0.85)
- Not checked — no full text of the sourcecites [26] [26] C.-L. Li, K. Sohn, J. Yoon, and T. Pfister. CutPaste: Self-supervised learning for anomaly detection and localization. In IEEE Conf. Comput. Vis. Pattern Recog. , 2021. - source
- Partly supportedcites [28] [28] P. Liznerski, L. Ruff, R. A. Vandermeulen, B. J. Franks, M. Kloft, and K.-R. Müller. Explainable deep one-class classification. In Int. Conf. Learn. Represent. , 2021. - sourceExperimental — accuracy not yet measured
Section 4.1 (p. 5) describes a common one-vs-rest anomaly-detection benchmark that treats each class in turn as nominal, consistent with training class-specific models. It does not establish that existing unsupervised anomaly-detection methods generally require a separate model for every object class. (LLM, conf=0.84)
- Not checked — no full text of the sourcecites [49] [49] J. Yi and S. Yoon. Patch SVDD: Patch-level SVDD for anomaly detection and segmentation. In Asian Conf. Comput. Vis. , 2020. - source
- Partly supportedcites [50] [50] Z. You, K. Yang, W. Luo, L. Cui, Y. Zheng, and X. Le. ADTR: Anomaly detection transformer with feature reconstruction. arXiv preprint arXiv:2209.01816 , 2022. - sourceExperimental — accuracy not yet measured
The Dataset section (p. 6) describes a one-class CIFAR-10 setup in which the training set of one class is used for training, supporting class-specific training in that setting. The paper does not establish that existing unsupervised anomaly detection methods generally require separate models for different objects. (LLM, conf=0.75)
- Not checked — no full text of the sourcecites [53] [53] V. Zavrtanik, M. Kristan, and D. Skoˇ caj. DRAEM-A discriminatively trained reconstruction embedding for surface anomaly detection. In Int. Conf. Comput. Vis. , 2021. - source
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. 1 · plain
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“EfficientNet performs obviously better than ResNet, especially in anomaly detection.”
What the sources say
- Not checked — no full text of the sourcecites [20] [20] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE Conf. Comput. Vis. Pattern Recog. , 2016. - source
- Partly supportedcites [41] [41] M. Tan and Q. Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Int. Conf. Mach. Learn. , 2019. - sourceExperimental — accuracy not yet measured
The Introduction (p. 2) reports higher ImageNet accuracy for EfficientNet-B4 than ResNet-50, and the transfer-learning discussion (p. 8) reports better accuracy–parameter performance than ResNet. The paper does not report an anomaly-detection comparison, so it does not support the claim that EfficientNet performs especially better in that task. (LLM, conf=0.90)
boostedstates an association- 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.
Every source the paper cites for this claim was checked and supports it.
In the paper
“Adding only one query embedding to the first decoder layer ( i.e. , vanilla transformer [43]) promotes the performance dramatically by 18.1% and 13.4% in anomaly detection and localization, respectively.”
What the sources say
- Supportedcites [43] [43] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need. In Adv. Neural Inform. Process. Syst. , 2017. - source
The citation is attached to “vanilla transformer,” identifying the architecture used as a baseline, rather than to the reported performance gains. The Introduction (p. 2) proposes the Transformer, and Section 3.1 (p. 3) describes its decoder layers. The 18.1% and 13.4% results are not attributed to this source by that citation. (LLM, conf=0.92)
boostedstates 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.
- Supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“Inspired by Denoising Auto-Encoder (DAE) [3, 44], we add perturbations to feature tokens, guiding the model to learn knowledge of normal samples by the denoising task.”
What the sources say
- Supportedcites [3] [3] Y. Bengio, L. Yao, G. Alain, and P. Vincent. Generalized denoising auto-encoders as generative models. In Adv. Neural Inform. Process. Syst. , 2013. - source
The Introduction describes corrupting observed inputs and training a denoising auto-encoder to reconstruct them. The Conclusion says denoising training implicitly estimates the underlying data-generating process. The citing authors present perturbing feature tokens as their own DAE-inspired implementation, not as a procedure used in this source. (LLM, conf=0.90)
- Not checked — no full text of the sourcecites [44] [44] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol. Extracting and composing robust features with denoising autoencoders. In Int. Conf. Mach. Learn. , 2008. - source
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.
- Rests on 2 assumptions not shown here.
- Supported
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
“NMA together with FJ provides a quite strong performance (96.2% for detection and 96.6% for localization).”
boostedstates 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.
72 more claims
Create a free account to see all 114 claims and to verify the 1 reference still open.
- It may state two separate claims as one. 25
- Probably not a substantial claim. 14
- The quote only partly states this claim. 2
Citations
Cites
The publisher hasn’t deposited this paper’s reference list.
Research works printed but not linked: 4
- [10][10] L. Deecke, L. Ruff, R. A. Vandermeulen, and H. Bilen. Transfer-based semantic anomaly detection. In Int. Conf. Mach. Learn. , 2021. -
- [22][22] L. Ilya and H. Frank. Decoupled weight decay regularization. In Int. Conf. Learn. Represent. , 2019. -
- [24][24] A. Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront , 2009. -
- [36][36] L. Ruff, R. Vandermeulen, N. Goernitz, L. Deecke, S. A. Siddiqui, A. Binder, E. Müller, and M. Kloft. Deep one-class classification. In Int. Conf. Mach. Learn. , 2018. -
Cited by
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
AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models
Gu, Z., Zhu, B., Zhu, G., Chen, Y., Tang, M., & Wang, J. (2024) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v38i3.27963 · Cited by 298
A Diffusion-Based Framework for Multi-Class Anomaly Detection
He, H., Zhang, J., Chen, H., Chen, X., Li, Z., Chen, X., Wang, Y., Wang, C., & Xie, L. (2024) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v38i8.28690 · Cited by 212
RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection
Zhang, X., Xu, M., & Zhou, X. (2024) · 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52733.2024.01580 · Cited by 209
Real-IAD: A Real-World Multi-View Dataset for Benchmarking Versatile Industrial Anomaly Detection
Wang, C., Zhu, W., Gao, B.-B., Gan, Z., Zhang, J., Gu, Z., Qian, S., Chen, M., & Ma, L. (2024) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr52733.2024.02159 · Cited by 353
Metadata
- Authors
- You, Z., Cui, L., Shen, Y., Yang, K., Lu, X., Zheng, Y., & Le, X.
- ORCID
- Lu, Xin
- Venue
- Advances in Neural Information Processing Systems 35
- Published
- 2022
- Pages
- 4571-4584
- ISSN
- 1049-5258 (ISSN-L)
- OpenAlex
- W7133227337
- Cited by
- 303
- References
- 0
- Citations by year
- 2026: 115 · 2025: 76 · 2024: 87 · 2023: 25
- FWCI
- 25.97
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- neural information processing systems (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- National Natural Science Foundation of China (62176152)
- National Key Research and Development Program of China (2021YFB1716000)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Anomaly Detection Techniques and Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Imbalanced Data Classification Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Adversarial Robustness in Machine Learning
- Keywords
- multi-class anomaly detection, unified model, anomaly detection
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.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
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: A Unified Model for Multi-Class Anomaly Detection. https://app.cochat.ai/papers/10.52202/068431-0330
- The paper
- You, Z., Cui, L., Shen, Y., Yang, K., Lu, X., Zheng, Y., & Le, X. (2022). A Unified Model for Multi-Class Anomaly Detection. Advances in Neural Information Processing Systems 35. https://doi.org/10.52202/068431-0330
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.