Pith. sign in

REVIEW 4 major objections 4 minor 38 references

KairosAD: A SAM-Based Model for Industrial Anomaly Detection on Embedded Devices

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read KairosAD claims that a frozen MobileSAM encoder plus a tiny fully connected head performs image-level industrial anomaly detection at 99.10% I-AUROC with 11.53M parameters and 5 ms inference, 78% fewer parameters than the most accurate…

desk verdict A small, useful edge-anomaly recipe whose headline accuracy number cannot be checked until the authors disclose where their positive training labels came from. read the letter →

arxiv 2505.24334 v1 pith:FKBDSLMY submitted 2025-05-30 cs.CV

Uzair Khan , Franco Fummi , Luigi Capogrosso This is my paper
classification cs.CV
keywords IndustrialanomalydetectionMobileSAMEmbeddedsystemsEfficientdeeplearningVisualfoundationmodelsMVTec-ADViSAImage-level
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

KairosAD is a supervised, whole-image anomaly detector aimed at industrial production lines with limited computing power. The paper claims that a MobileSAM image encoder, used exactly as pretrained, gives embeddings from which a few fully connected layers can separate normal from anomalous product images: on MVTec-AD the model reaches 99.10% I-AUROC with 11.53 million parameters and a 5 ms GPU inference time. That is 78% fewer parameters than SimpleNet, the most accurate baseline, and a 4x faster inference than STLM, the leading efficient competitor, with a slightly higher I-AUROC than STLM (99.10% versus 99.05%). The same model ran on NVIDIA Jetson NX and AGX and on a real production line at the University of Verona, which is the practical point: accuracy close to much larger models at a size and speed that fit embedded devices. On ViSA the average I-AUROC drops to 92.3%, fourth among the compared methods, and the paper flags the Capsules class as its main weakness.

What carries the argument

The load-bearing component is the MobileSAM image encoder, a lightweight distilled version of SAM's ViT-based encoder, used frozen exactly as pretrained: the paper states that KairosAD 'utilizes the MobileSAM image encoder as it is' (Eq. 1). On top of the resulting image embedding sits the Anomaly Score Predictor (ASP), a stack of two or three fully connected layers with ReLU, mapping the embedding to a score in [0,1] trained with a weighted binary cross-entropy loss (Eq. 3) that reweights the positive class by the negative-to-positive sample ratio. The entire system is a single-branch architecture, which the paper contrasts with STLM's two-branch design as the source of its speed and simplicity; the frozen encoder supplies all representational power, and the small head learns the normal-versus-anomalous separation.

What would settle it

Run the exact KairosAD training protocol (two fully connected layers, 35 epochs, Adam at 1e-2, weighted BCE) on MVTec-AD but swap the frozen MobileSAM encoder for another frozen lightweight backbone of comparable size, such as ResNet-18 or DINOv2-small. If that model also reaches roughly 99% I-AUROC, then MobileSAM's specific features are not what carries the result, and the paper's explanation for the efficiency-accuracy balance fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single-branch design—the MobileSAM image encoder kept frozen and used as-is, followed by two (MVTec-AD) or three (ViSA) fully connected layers with ReLU, trained with weighted binary cross-entropy—performs image-level anomaly detection at accuracy comparable to much larger state-of-the-art models. On MVTec-AD, KairosAD reports 99.10% I-AUROC with 11.53M parameters and 5 ms latency; the paper contrasts this with STLM (99.05%, 16.56M, 20 ms) and SimpleNet (99.57%, 52.88M, 39 ms), framing the result as 78% fewer parameters than SimpleNet and a 4x speedup over STLM. On ViSA it reports 92.3% average I-AUROC with 10.55M parameters and 5 ms, ranking fourth in accuracy but first in speed and parameter efficiency. The paper also reports deployment on NVIDIA Jetson NX (211 ms) and AGX (218 ms), a live installation on the production line of the ICE Lab, and training energy of 0.0465 kWh with 0.0154 kg CO2e emissions.

Load-bearing premise

The load-bearing premise is that the frozen MobileSAM encoder, trained on natural-image segmentation, produces embeddings in which normal and anomalous industrial product images are cleanly separable for every product category, so a two- or three-layer classifier reaches about 99% accuracy without fine-tuning or domain adaptation.

Editorial extensions

If this is right

  • On MVTec-AD, the efficiency-accuracy trade-off improves to 99.10% I-AUROC at 11.53M parameters and 5 ms, so image-level anomaly detection no longer requires memory-bank or reconstruction models that are an order of magnitude larger.
  • KairosAD dominates STLM on all three reported MVTec-AD metrics (accuracy 99.10% vs 99.05%, parameters 11.53M vs 16.56M, latency 5 ms vs 20 ms), which would make STLM obsolete as the efficient baseline.
  • The 78% parameter cut relative to SimpleNet (11.53M vs 52.88M) comes at a 0.47-point I-AUROC cost (99.10 vs 99.57) but buys a greater than 7x speedup, a trade favorable for real-time embedded use.
  • On ViSA, ranking fourth in average I-AUROC (92.3%) while being the fastest (5 ms) and smallest (10.55M) model shows the efficiency advantage is not specific to MVTec-AD, though the Capsules class (69.7% I-AUROC) is a clear failure case.
  • Deployment on Jetson NX (211 ms) and Jetson AGX (218 ms) and on the ICE Lab production line indicates the model can operate outside the GPU benchmark setting, on hardware small manufacturers actually use.

Reading between the lines

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

  • Because the paper never replaces MobileSAM with another frozen backbone, it cannot isolate how much of the accuracy comes from the encoder's features versus the supervised head; a backbone-swap experiment would settle which ingredient matters.
  • The uneven per-class results on ViSA's Capsules and MVTec-AD's Metal nut and Wood suggest the frozen encoder's features are stronger for some defect types; a lightly fine-tuned or domain-adapted encoder is the obvious next experiment to close those gaps.
  • Because the pipeline is supervised, its embedded-deployment advantage does not automatically carry to the unsupervised anomaly detection setting that dominates industrial practice; extending the recipe with few-shot or self-supervised heads is a natural next test.
  • The 5 ms figure is GPU-only, while the reported Jetson latencies of 211–218 ms imply roughly five images per second on edge hardware, so production-line evaluation should focus on end-to-end edge throughput rather than the GPU benchmark.
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 / 4 minor

Summary. The paper proposes KairosAD, an image-level industrial anomaly detector that couples a frozen MobileSAM image encoder with a small fully connected head (the ASP sub-network) trained by weighted binary cross-entropy. The authors report 99.10% I-AUROC on MVTec-AD and 92.3% on ViSA, with 11.53M parameters and 5 ms inference time on an RTX 4090, and they describe a deployment on Jetson NX/AGX and on a real production line. The central claims are that KairosAD requires 78% fewer parameters than SimpleNet and is 4x faster than STLM while maintaining comparable AUROC, making it suitable for embedded industrial use.

Significance. If the results are valid, the paper would make a useful practical contribution: it demonstrates that a foundation-model encoder plus a tiny classification head can run at high speed on low-power hardware while retaining competitive image-level anomaly detection accuracy. The deployment on actual embedded devices and a production line strengthens the application value. However, the validity of the performance claims is currently undermined by an underspecified and potentially invalid training protocol, because the WBCE loss in Eq. (3) requires anomalous labels while the MVTec-AD and ViSA official training splits contain only normal images. Until this is resolved, the reported I-AUROC numbers cannot be trusted as a fair comparison against the unsupervised baselines. The paper also contains several internal inconsistencies (tables, formal definitions, latency figures) that must be corrected.

major comments (4)
  1. [Section 4.1] The WBCE loss in Eq. (3) uses binary labels y_i with both normal (0) and anomalous (1) classes, and the implementation details state that the model is trained for 35 epochs on MVTec-AD and 50 epochs on ViSA. However, the official MVTec-AD and ViSA training splits contain only normal images; anomalous images appear only in the test splits. The manuscript does not describe any custom train/test split, any held-out anomalous training images, or any other source of positive labels. As written, the reported training procedure is impossible to execute on the official training data, and the only apparent way to obtain positive labels is to use test-set images, which would constitute label leakage and invalidate the reported I-AUROC of 99.10% on MVTec-AD and 92.3% on ViSA. The authors must specify exactly which images and labels were used for training, and they must demonstrate that no test images were seen during training. This is the load-bearing point for the paper's central claim of maintaining comparable AUROC.
  2. [Section 3.1] The formal definition states that the image embedding E has dimension d = 3. This is incorrect for the MobileSAM image encoder, which outputs a high-dimensional feature map (typically 256 channels per spatial token, or a global embedding of much higher dimension). A d = 3 output is not consistent with the described architecture, and it makes the subsequent parameter counts (11.53M total) impossible to verify: a 3-dimensional input to the ASP head would require only a negligible number of parameters. The authors should correct the embedding dimension and provide the actual output shape of the encoder, as the current description prevents any reproduction of the architecture.
  3. [Tables 1 and 2] The latency and parameter values for PatchCore and RD4AD are inconsistent between the two tables. In Table 1, PatchCore is reported with latency 180 ms and 186.55M parameters, and RD4AD with latency 28 ms and 150.64M parameters. In Table 2, PatchCore appears with latency 37 ms and 94.70M parameters, and RD4AD with latency 180 ms and 186.55M parameters. These numbers appear swapped or misattributed relative to Table 1. Since the paper uses these numbers to support its efficiency claims, the discrepancies undermine the reliability of the comparison and must be reconciled.
  4. [Section 4.4] Section 4.4 reports inference times of 211 ms and 218 ms on the NVIDIA Jetson NX and AGX, respectively, whereas Table 1 reports a 5 ms inference time for KairosAD on the MVTec-AD benchmark. The manuscript does not specify the hardware platform for Table 1's timings. Because the abstract and Section 4.2 claim a '4x faster inference time' compared to STLM, it is essential to state clearly which platform the 5 ms figure was measured on and how the embedded-deployment timings relate to the headline speed claim.
minor comments (4)
  1. [Figure 2 caption] The word 'architetcure' is a typo and should be 'architecture'.
  2. [Table 1 caption] The phrase 'In gold, the best results' likely refers to bold or highlighted values; this should be clarified and the table formatting should be consistent.
  3. [Section 4.4 and Figure 3] The qualitative results in Figure 3 are hard to interpret without a scale bar or a clearer description of the production-line setup; please add details about the camera, lighting, and product types.
  4. [General] The paper compares a supervised method (KairosAD) against unsupervised baselines without discussing whether this is a fair comparison or whether the baselines were given any labeled information; this should be addressed in the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: KairosAD's accuracy, latency, and parameter-count claims are measured quantities or additive checkpoint counts, with no load-bearing step that reduces to its own inputs.

full rationale

The central claims (99.10% I-AUROC on MVTec-AD, 11.53M parameters, 5 ms inference, 78% parameter reduction vs. SimpleNet, 4x speedup vs. STLM) are empirical measurements on held-out test data and direct model/checkpoint counts, not quantities derived from the assumptions that define the model. The architecture is a frozen MobileSAM encoder (Eq. 1) followed by a small MLP trained with weighted binary cross-entropy (Eqs. 2-3). Nothing in the training objective or architecture is defined in terms of the reported AUROC or latency values; the evaluation is an external comparison against unrelated baselines. The MobileSAM encoder is an externally pretrained, frozen backbone, so even the strongest underlying assumption—that its embeddings separate normal from anomalous industrial images—is empirical and testable rather than definitionally guaranteed. Self-citations (refs. 5, 6, 11, 16) appear only in contextual statements about tiny ML and predictive maintenance and do not carry the accuracy or efficiency claims. The paper's use of the term 'supervised' with a loss that requires anomalous labels, while MVTec-AD's official training split contains only normal images, is a potentially serious experimental-protocol concern that could undermine the accuracy claim if positive labels were taken from the test set. However, that issue is about data leakage and missing experimental detail, not about circularity: a leaked or incorrectly obtained label is an input artifact, not a derivation that is equivalent by construction to the claimed result. Therefore, per the rubric, the honest finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the transferability of frozen MobileSAM features, the correctness of the datasets and competitor numbers, and the availability of supervised labels; these are all domain assumptions without independent verification in this paper. The free parameters are the MLP weights and a few hyperparameters; there are no created entities.

free parameters (5)
  • ASP fully connected weights (phi) = 11.53M total on MVTec-AD; 10.55M on ViSA
    The MLP head is trained on each dataset with the WBCE loss; these are the model parameters being fitted.
  • Number of fully connected layers N_l = 2 on MVTec-AD, 3 on ViSA
    Chosen per dataset without ablation or justification, introducing a dataset-specific capacity adjustment.
  • WBCE class weight w_i = Ratio of negative to positive samples
    Standard inverse-frequency weighting, hand-set rather than learned.
  • ASP input embedding dimension d = 3 as written in Eq. (1)
    Almost certainly a typo; the actual input dimension of the MLP is never specified, a missing implementation detail.
  • Learning rate and training epochs = 1e-2; 35 epochs on MVTec-AD, 50 on ViSA
    Chosen without reported sensitivity analysis; the different epoch counts per dataset are another dataset-specific choice.
assumptions (3)
  • domain assumption MobileSAM's frozen pretrained image encoder transfers to industrial defect detection.
    Section 3.1: the paper 'utilizes the MobileSAM image encoder as it is'. No fine-tuning or domain adaptation is done, so the model's accuracy depends on this transfer holding.
  • domain assumption The MVTec-AD and ViSA test labels and the competitor results taken from prior papers are accurate.
    Tables 1 and 2 compile competitor numbers from [19] and [34] without reproducing them; the two tables disagree on PatchCore and RD4AD, so this assumption is partially violated.
  • domain assumption Supervised labels for both normal and anomalous samples are available during training.
    The training loss in Eq. (3) is supervised WBCE with labels yi. Many industrial anomaly detection scenarios provide only normal samples; the paper does not discuss this constraint.

how reviews work

0 comments
Cite this review

Pith. "Pith review of KairosAD: A SAM-Based Model for Industrial Anomaly Detection on Embedded Devices." pith.science (2026). https://pith.science/paper/FKBDSLMY

@misc{pith2026250524334,
  author       = {Pith},
  title        = {Pith review of: KairosAD: A SAM-Based Model for Industrial Anomaly Detection on Embedded Devices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FKBDSLMY}},
  note         = {Machine review of arXiv:2505.24334}
}
read the original abstract

In the era of intelligent manufacturing, anomaly detection has become essential for maintaining quality control on modern production lines. However, while many existing models show promising performance, they are often too large, computationally demanding, and impractical to deploy on resource-constrained embedded devices that can be easily installed on the production lines of Small and Medium Enterprises (SMEs). To bridge this gap, we present KairosAD, a novel supervised approach that uses the power of the Mobile Segment Anything Model (MobileSAM) for image-based anomaly detection. KairosAD has been evaluated on the two well-known industrial anomaly detection datasets, i.e., MVTec-AD and ViSA. The results show that KairosAD requires 78% fewer parameters and boasts a 4x faster inference time compared to the leading state-of-the-art model, while maintaining comparable AUROC performance. We deployed KairosAD on two embedded devices, the NVIDIA Jetson NX, and the NVIDIA Jetson AGX. Finally, KairosAD was successfully installed and tested on the real production line of the Industrial Computer Engineering Laboratory (ICE Lab) at the University of Verona. The code is available at https://github.com/intelligolabs/KairosAD.

Figures

Figures reproduced from arXiv: 2505.24334 by the authors.

Figure 1
Figure 1. Comparisons of different anomaly detection methods in terms of I-AUROC (vertical axis), inference time (horizontal axis), and the ratios of parameter numbers (circle radius), on the MVTec-AD dataset [2]. Our KairosAD achieves competitive I-AUROC results, while having 78% fewer parameters with respect to SimpleNet [24], i.e., the model that achieves the highest AUROC. Instead, with respect to STLM [19], which represe… view at source ↗
Figure 2
Figure 2. The KairosAD architetcure. Specifically, KairosAD is composed of a Seman￾tic Feature Encoder (SFE) and an Anomaly Score Prediction (ASP). This simple but effective model allows us to perform efficient image-level anomaly detection on embed￾ded devices. models capable of delivering acceptable performances for their intended appli￾cations, specialized techniques have been proposed in model architecture ex￾ploration, m… view at source ↗
Figure 3
Figure 3. Qualitative results of KairosAD in correctly detecting (and not detecting anomalies) on the MVTec-AD, ViSA, and the real production line of the ICE Lab at the University of Verona. For each image, we report the [True Label] - [Prediction] on the top left (where 0 stands for normal and 1 stands for abnormal). The anomalies are circled in blue. 5 Conclusions In this article, we present KairosAD, an efficient image-bas… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 36 canonical work pages

  1. [1]

    In: Asian Conference on Computer Vision (ACCV) (2019)

    Akcay, S., Atapour-Abarghouei, A., Breckon, T.P.: GANomaly: Semi-supervised Anomaly Detection via Adversarial Training. In: Asian Conference on Computer Vision (ACCV) (2019)

    work page 2019

  2. [2]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2019)

    Bergmann, P., Fauser, M., Sattlegger, D., Steger, C.: MVTec AD — A Compre- hensive Real-World Dataset for Unsupervised Anomaly Detection. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2019)

    work page 2019

  3. [3]

    In: Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (2019)

    Bergmann, P., Löwe, S., Fauser, M., Sattlegger, D., Steger, C.: Improving Unsu- pervised Defect Segmentation by Applying Structural Similarity to Autoencoders. In: Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (2019)

    work page 2019

  4. [4]

    IEEE Transactions on Knowledge and Data Engineering35(4), 3279– 3298 (2023)

    Brauwers, G., Frasincar, F.: A General Survey on Attention Mechanisms in Deep Learning. IEEE Transactions on Knowledge and Data Engineering35(4), 3279– 3298 (2023)

    work page 2023

  5. [5]

    IEEE Access 12, 23406– 23426 (2024) KairosAD: Industrial Anomaly Detection on Embedded Devices 11

    Capogrosso, L., Cunico, F., Cheng, D.S., Fummi, F., Cristani, M.: A Machine Learning-Oriented Survey on Tiny Machine Learning. IEEE Access 12, 23406– 23426 (2024) KairosAD: Industrial Anomaly Detection on Embedded Devices 11

    work page 2024

  6. [6]

    In: 19th International Conference on Computer Vision Theory and Applications (VISAPP) (2024)

    Capogrosso, L., Girella, F., Taioli, F., Chiara, M., Aqeel, M., Fummi, F., Setti, F., Cristani, M.: Diffusion-Based Image Generation for In-Distribution Data Augmen- tation in Surface Defect Detection. In: 19th International Conference on Computer Vision Theory and Applications (VISAPP) (2024)

    work page 2024

  7. [7]

    In: International Conference on Computer Vision (ICCV) (2021)

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging Properties in Self-Supervised Vision Transformers. In: International Conference on Computer Vision (ICCV) (2021)

    work page 2021

  8. [8]

    In: Winter Conference on Applications of Computer Vision (WACV) (2024)

    Damm, S., Laszkiewicz, M., Lederer, J., Fischer, A.: AnomalyDINO: Boosting Patch-based Few-shot Anomaly Detection with DINOv2. In: Winter Conference on Applications of Computer Vision (WACV) (2024)

    work page 2024

Show all 38 references
  1. [9]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    Deng, H., Li, X.: Anomaly Detection via Reverse Distillation from One-Class Em- bedding. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    2022

  2. [10]

    Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M.W., Keutzer, K.: A Sur- vey of Quantization Methods for Efficient Neural Network Inference, pp. 291–326. Chapman and Hall/CRC (2022)

    2022

  3. [11]

    In: International Conference on Content-Based Multi- media Indexing (CBMI) (2024)

    Girella, F., Liu, Z., Fummi, F., Setti, F., Cristani, M., Capogrosso, L.: Leveraging Latent Diffusion Models for Training-Free in-Distribution Data Augmentation for Surface Defect Detection. In: International Conference on Content-Based Multi- media Indexing (CBMI) (2024)

    2024

  4. [12]

    Inter- national Journal of Computer Vision129(6), 1789–1819 (2021)

    Gou, J., Yu, B., Maybank, S.J., Tao, D.: Knowledge Distillation: A Survey. Inter- national Journal of Computer Vision129(6), 1789–1819 (2021)

    2021

  5. [13]

    In: Win- ter Conference on Applications of Computer Vision (WACV) (2022)

    Gudovskiy, D., Ishizaka, S., Kozuka, K.: CFLOW-AD: Real-Time Unsupervised Anomaly Detection with Localization via Conditional Normalizing Flows. In: Win- ter Conference on Applications of Computer Vision (WACV) (2022)

    2022

  6. [14]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2022)

    Jiang, X., Liu, J., Wang, J., Nie, Q., Wu, K., Liu, Y., Wang, C., Zheng, F.: Soft- Patch: Unsupervised Anomaly Detection with Noisy Data. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)

    2022

  7. [15]

    ACM Computing Surveys54(10), 1–41 (2022)

    Khan, S., Naseer, M., Hayat, M., Zamir, S.W., Khan, F.S., Shah, M.: Transformers in Vision: A Survey. ACM Computing Surveys54(10), 1–41 (2022)

    2022

  8. [16]

    ACM Trans- actions on Embedded Computing Systems (2025)

    Khan, U., Cheng, D., Setti, F., Fummi, F., Cristani, M., Capogrosso, L.: A Com- prehensive Survey on Deep Learning-based Predictive Maintenance. ACM Trans- actions on Embedded Computing Systems (2025)

    2025

  9. [17]

    In: Interna- tional Conference for Learning Representations (ICLR) (2014)

    Kingma, D.P., Ba, J.: Adam: A Method for Stochastic Optimization. In: Interna- tional Conference for Learning Representations (ICLR) (2014)

    2014

  10. [18]

    In: International Conference on Computer Vision (ICCV) (2023)

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment Anything. In: International Conference on Computer Vision (ICCV) (2023)

    2023

  11. [19]

    ACM Transactions on Multimedia Computing, Communica- tions, and Applications21(2), 1–23 (2025)

    Li, C., Qi, L., Geng, X.: A SAM-guided Two-stream Lightweight Model for Anomaly Detection. ACM Transactions on Multimedia Computing, Communica- tions, and Applications21(2), 1–23 (2025)

    2025

  12. [20]

    Foundations and Trends in Computer Graphics and Vision16(1–2), 1–214 (2024)

    Li, C., Gan, Z., Yang, Z., Yang, J., Li, L., Wang, L., Gao, J.: Multimodal Foun- dation Models: From Specialists to General-Purpose Assistants. Foundations and Trends in Computer Graphics and Vision16(1–2), 1–214 (2024)

    2024

  13. [21]

    IEEE Transac- tions on Image Processing32, 4327–4340 (2023)

    Liang, Y., Zhang, J., Zhao, S., Wu, R., Liu, Y., Pan, S.: Omni-Frequency Channel- Selection Representations for Unsupervised Anomaly Detection. IEEE Transac- tions on Image Processing32, 4327–4340 (2023)

    2023

  14. [22]

    Machine Intelligence Research21(1), 104– 135 (2024) 12 U

    Liu, J., Xie, G., Wang, J., Li, S., Wang, C., Zheng, F., Jin, Y.: Deep Industrial Image Anomaly Detection: A Survey. Machine Intelligence Research21(1), 104– 135 (2024) 12 U. Khan et al

    2024

  15. [23]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

    Liu, W., Li, R., Zheng, M., Karanam, S., Wu, Z., Bhanu, B., Radke, R.J., Camps, O.: Towards Visually Explaining Variational Autoencoders. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2020)

    2020

  16. [24]

    In: Conference on Computer Vision and Pat- tern Recognition (CVPR) (2023)

    Liu, Z., Zhou, Y., Xu, Y., Wang, Z.: SimpleNet: A Simple Network for Image Anomaly Detection and Localization. In: Conference on Computer Vision and Pat- tern Recognition (CVPR) (2023)

    2023

  17. [25]

    Transactions on Machine Learning Research (2024)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., HAZIZA, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P.Y., Li, S.W., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jegou, H., Mairal, ...

    2024

  18. [26]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2019)

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. In: Advances in Neural Information Processing Systems (NeurIPS) (2019)

    2019

  19. [27]

    In: 38th International Conference on Machine Learning (ICML) (2021)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning Trans- ferable Visual Models From Natural Language Supervision. In: 38th International Conference on Machine Learning (IC...

    2021

  20. [28]

    ACM Computing Surveys 54(4), 1–34 (2021)

    Ren, P., Xiao, Y., Chang, X., Huang, P.y., Li, Z., Chen, X., Wang, X.: A Com- prehensive Survey of Neural Architecture Search: Challenges and Solutions. ACM Computing Surveys 54(4), 1–34 (2021)

    2021

  21. [29]

    In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    Roth, K., Pemula, L., Zepeda, J., Scholkopf, B., Brox, T., Gehler, P.: Towards Total Recall in Industrial Anomaly Detection. In: Conference on Computer Vision and Pattern Recognition (CVPR) (2022)

    2022

  22. [30]

    In: Con- ference on Computer Vision and Pattern Recognition (CVPR) (2023)

    Tien, T.D., Nguyen, A.T., Tran, N.H., Huy, T.D., Duong, S.T., Nguyen, C.D.T., Truong, S.Q.H.: Revisiting Reverse Distillation for Anomaly Detection. In: Con- ference on Computer Vision and Pattern Recognition (CVPR) (2023)

    2023

  23. [31]

    IEEE Access 10, 63280–63300 (2022)

    Vadera, S., Ameen, S.: Methods for Pruning Deep Neural Networks. IEEE Access 10, 63280–63300 (2022)

    2022

  24. [32]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention Is All You Need. In: Advances in Neural Information Processing Systems (NeurIPS) (2017)

    2017

  25. [33]

    International Journal of Computer Vision132(12), 5635–5662 (2024)

    Yang, J., Zhou, K., Li, Y., Liu, Z.: Generalized Out-of-Distribution Detection: A Survey. International Journal of Computer Vision132(12), 5635–5662 (2024)

    2024

  26. [34]

    In: European Conference on Computer Vision (ECCV) (2024)

    Yao, H., Liu, M., Yin, Z., Yan, Z., Hong, X., Zuo, W.: GLAD: Towards Better Reconstruction with Global and Local Adaptive Diffusion Models for Unsupervised Anomaly Detection. In: European Conference on Computer Vision (ECCV) (2024)

    2024

  27. [35]

    In: International Con- ference on Computer Vision (ICCV) (2023)

    Yao, X., Li, R., Qian, Z., Luo, Y., Zhang, C.: Focus the Discrepancy: Intra- and Inter-Correlation Learning for Image Anomaly Detection. In: International Con- ference on Computer Vision (ICCV) (2023)

    2023

  28. [36]

    In: International Con- ference on Computer Vision (ICCV) (2021)

    Zavrtanik, V., Kristan, M., Skocaj, D.: DRAEM - A Discriminatively Trained Reconstruction Embedding for Surface Anomaly Detection. In: International Con- ference on Computer Vision (ICCV) (2021)

    2021

  29. [37]

    arXiv preprint arXiv:2306.14289 (2023)

    Zhang, C., Han, D., Qiao, Y., Kim, J.U., Bae, S.H., Lee, S., Hong, C.S.: Faster Seg- ment Anything: Towards Lightweight SAM for Mobile Applications. arXiv preprint arXiv:2306.14289 (2023)

    2023 arXiv

  30. [38]

    In: European Conference on Computer Vision (ECCV) (2022)

    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 (ECCV) (2022)

    2022

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.