Pith. sign in

REVIEW 3 major objections 4 minor 43 references

IMMoE: Incomplete Multi-View Anomaly Detection via Mixture of View Experts Fusion

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper argues that incomplete multi-view anomaly detection is best handled by fusing view-specific experts, and it reports that its reconstruction model outperforms existing methods on both a new masked benchmark and a public multi-view

desk verdict A plausible engineering combination and a useful new benchmark, but the test-time access to the unmasked original image for computing F_gt is ambiguous and could invalidate the central claim. read the letter →

arxiv 2607.19032 v1 pith:WOHSPGSS submitted 2026-07-21 cs.CV

Lei Hu This is my paper
classification cs.CV
keywords incompletemulti-viewanomalydetectionindustrialfusionmixtureofexpertsocclusionrobustnessreconstruction-basedbenchmarkdatasetfeaturedropout
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

The paper argues that real industrial inspection must cope with incomplete views, where occlusion or sensor faults hide parts of a product. It introduces a new task, incomplete multi-view anomaly detection, and a benchmark built by masking 50% of each product area. Its method, IMMoE, trains a reconstruction model to borrow missing information from other views through a gated mixture of view-specific experts, while a dropout-based encoder prevents the model from memorizing masked regions. The authors report that IMMoE outperforms prior methods on their benchmark and also on a complete multi-view dataset, suggesting the fusion mechanism helps even when views are intact.

What carries the argument

Multi-View Expert Fusion (MVEF) is the central mechanism: a set of view-specific linear experts, a gating network that scores each view's contribution, and cross-attention that folds the weighted global representation back into each view so reconstruction can borrow missing content. Local Anomaly Enhancement Encoder (LAEE) applies dropout to patch features followed by linear attention, stopping the model from treating masks as normal texture. Area Adaptive Loss weights gradients by local reconstruction difficulty and region priors, prioritizing masked regions and product foregrounds over background.

What would settle it

Take a set of real defective products photographed under natural occlusion—part of the object covered by packaging, hands, or machine parts—with pixel-level ground truth, and run IMMoE against a strong single-view baseline. If IMMoE's lead shrinks to near zero or reverses on such data, the central claim that expert fusion restores missing view information would be falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that cross-view fusion, not stronger single-view modeling, is the key to anomaly detection under missing information. In IMMoE, each view is processed by its own expert; a gating network weights the experts to form a global fused representation, and cross-attention injects that global context back into each view's reconstruction. A separate encoder applies dropout to local features so the network cannot overfit the mask, and a region-adaptive loss up-weights incomplete product areas. On the new RIMAD benchmark, the paper reports 84.7% image-level AUROC and 33.1% pixel-level AP, exceeding the best previous method by 2.8 and 11.8 points; on the complete multi-view

Load-bearing premise

The whole evaluation rests on the assumption that randomly masking 50% of the segmented product area with small rectangles reproduces the way occlusions actually happen on real production lines; if it does not, the benchmark's ranking and the model's advantage may not transfer to genuine occluded data.

Editorial extensions

If this is right

  • Production-line inspection can tolerate partial occlusions or intermittent camera failures without retraining, since the model reconstructs hidden regions from other viewpoints.
  • The new masked benchmark gives the field a common test for robustness to missing view information, enabling direct comparison of future methods.
  • Because IMMoE also improves on complete-view data, view-expert fusion appears to be a general improvement, not just a repair for missing inputs.
  • Feature dropout as an anomaly regularizer may transfer to other reconstruction-based detectors, offering a simple defense against mask overfitting.

Reading between the lines

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

  • The synthetic masking uses random small rectangles; real occlusions are often large and structured, so the benchmark may not capture the hardest real cases. A real-occlusion test set would be the natural next check.
  • The view-expert fusion idea could apply to other incomplete-modality settings, such as multi-sensor or multi-camera inputs where an entire stream drops out, not just pixel masks.
  • The mask-ratio ablation shows performance rising again at 75% masking; one interpretation is that at high masking the model shifts toward pure reconstruction, which could obscure anomaly cues. This suggests testing whether high mask ratios hide defects rather than reveal them.
  • The method's dependence on a pretrained encoder and multiple expert networks raises a deployment question not addressed in the paper: a lightweight version would be needed for cost-sensitive production-line hardware.
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

3 major / 4 minor

Summary. The paper introduces Incomplete Multi-View Anomaly Detection (IMVAD), a task in which each test view is partially masked, and proposes IMMoE, a reconstruction-based framework with two main modules: a Local Anomaly Enhancement Encoder (LAEE) that applies dropout to local features, and a Multi-View Expert Fusion (MVEF) module that aggregates cross-view information with a mixture-of-experts gating mechanism. The authors also construct RIMAD by masking 50% of the SAM-segmented product area in each view of Real-IAD. They report state-of-the-art results on RIMAD (84.7% I-AUROC, 33.1% P-AP) and Real-IAD (89.7% I-AUROC, 44.3% P-AP), with ablations and qualitative visualizations.

Significance. If the empirical claims are sound, the paper makes a useful contribution: it proposes a new task variant, a benchmark generation pipeline, and a strong baseline that outperforms existing single-view and multi-view methods under incomplete inputs. The evaluation is broad in scope (30 categories, multiple recent baselines, ablation studies), and the authors promise to release code, which strengthens reproducibility. The main value would be as a benchmark and a practical method for industrial inspection under occlusion. However, the central claim depends on the exact inference protocol, which is currently ambiguous, and the reported margins on Real-IAD are small enough that statistical uncertainty may affect the comparison.

major comments (3)
  1. [§4.1, Fig. 3, §3.1] The test-time computation of F_gt is load-bearing and currently ambiguous. Section 4.1 states that a pre-trained encoder extracts F_in and F_gt from the incomplete and original images, respectively, and that in the inference phase the anomaly map is derived from the discrepancy between F_rec and F_gt. Under the task definition in Section 3.1, the test-time input is only the incomplete image x_i and its binary mask m_i; the original unmasked view is not available. If F_gt is obtained by passing the original image through the encoder, the reported RIMAD results use information that the IMVAD setting assumes is missing, which would give IMMoE an unfair advantage over baselines that receive only incomplete views. If F_gt is instead obtained from the incomplete input, the text and Figure 3 are misleading, and the loss/anomaly-score definition must be corrected. Please specify the exact infere
  2. [§5.1, Tables 1–5] The SOTA claim rests on single-run comparisons with no uncertainty estimates. Table 2 shows a 0.4 I-AUROC margin over Dinomaly on Real-IAD, which may lie within run-to-run noise; the paper reports no standard deviations or multiple seeds. Hyperparameters such as the dropout ratio p (Table 5) and the loss parameters θ and β (§5.1) are selected on the same benchmark and then used for the reported final numbers, with no explicit validation protocol. Report mean±std over at least three seeds, or provide confidence intervals, and state explicitly which hyperparameters were tuned on a validation split and whether the final test numbers reflect that selection.
  3. [§3.2, Table 4] The realism of the RIMAD masking simulation is not validated. The pipeline randomly discards small rectangles within the SAM-segmented product area until 50% is obscured, but real-world occlusions may be structured differently (large contiguous coverage, viewpoint-dependent shadows, complete view loss). Table 4 shows non-monotonic behavior across mask ratios and the paper provides no experiments with real occlusion data. Since RIMAD is the primary benchmark for the new task, this modeling assumption should be tested, or at least discussed with concrete failure cases; otherwise the transferability of the reported gains to actual industrial scenarios is unclear.
minor comments (4)
  1. [Table 2 vs Table 8] The average Real-IAD P-AP is 44.3 in Table 2 but 44.4 in Table 8. Please reconcile these numbers and audit all aggregate rows, including Table 1, against the per-category values.
  2. [§4.3, Eqs. (4)–(7)] The tensor dimensions in the MVEF module are under-specified. Fgate is defined in R^{V×N}, while each expert output is in R^{N×C}, so the multiplication in Eq. (7) needs a clear statement of how the view and patch dimensions are contracted or broadcast. Also state explicitly the softmax dimension in Eq. (5).
  3. [§4.4, Eqs. (11)–(12)] The text says the loss 'modulates the gradient' of each patch, but Eq. (11) defines a weighted loss value. Clarify whether w(j) directly scales the loss or is applied through a stop-gradient/backprop reweighting, and define β in the equation relative to the region prior values given in the text.
  4. [Fig. 4 and Fig. 6] The heatmaps are compared without a common color scale or colorbar, which limits the informativeness of the qualitative evaluation. Please include a shared colorbar or report quantitative localization maps.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's SOTA claims rest on standard empirical evaluation, not on a derivation that reduces to its inputs.

full rationale

IMMoE is an empirical anomaly-detection paper whose central claims are benchmark scores (Tables 1–2) obtained by training on normal samples and evaluating anomaly maps against ground-truth masks on held-out test data. No theoretical prediction is derived from a fitted parameter: the Area Adaptive Loss, dropout, and MVEF modules are architectural/training choices, and the reported numbers are comparisons, not quantities forced by the loss definition. The method builds on prior work such as Dinomaly and INP-Former, but those citations are not by the present author and are not used to justify an unverified equivalence or uniqueness result. The author's self-citations (e.g., ReplayCAD, table-structure recognition) appear only in related-work context and carry no load-bearing weight. The only notable concern is Section 4.1's definition of F_gt as features of 'original images' and the statement that the inference-phase anomaly map comes from the discrepancy between F_rec and F_gt; since the IMVAD problem definition supplies only x_i and m_i at test time, this is an evaluation-protocol ambiguity or potential leakage, not circularity. If the released code uses the unmasked image at inference, that would be a benchmark-validity flaw rather than a self-referential derivation. The paper makes no derivation that is equivalent to its inputs by construction, so no circular step is identified.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper rests on common empirical assumptions in reconstruction-based anomaly detection (frozen pretrained features, cosine-distance scoring) plus the new task-specific assumption that random rectangular masking approximates real occlusion. The main free parameters (dropout ratio, loss exponent, region priors, mask ratio) are tuned on the target datasets and are part of the reported recipe, but they do not enter a derivation. No new physical or conceptual entities are invented.

free parameters (4)
  • Dropout ratio p in LAEE = 0.1
    Tuned via ablation in Table 5; default 0.1 gives best I-AUROC.
  • Loss exponent θ in Area Adaptive Loss = 3
    Set by hand (Section 5.1); controls how much hard patches are amplified.
  • Region prior weights β = 0.4 (incomplete), 0.2 (foreground), 0 (background)
    Hand-chosen in Section 4.4 to emphasize masked product regions.
  • Mask ratio for RIMAD = 50%
    Dataset construction parameter chosen as the default (Section 3.2); ablations show non-monotonic sensitivity.
assumptions (5)
  • domain assumption Pre-trained DINOv2 encoder features provide a reliable basis for anomaly detection
    Used as the frozen feature extractor (Section 4.1); the whole reconstruction pipeline assumes these features carry the semantic content needed to localize defects.
  • domain assumption Cosine distance between reconstructed and input features is a valid anomaly score
    Inherited from INP-Former and Dinomaly (Section 4.1); no analysis of this metric's behavior is provided.
  • domain assumption Missing content in one view is recoverable from other views
    This is the core premise of the multi-view fusion module (Section 4.3); if views do not provide complementary information, the fusion cannot help.
  • domain assumption SAM segmentation masks are accurate enough to define product areas for masking
    The whole RIMAD construction depends on SAM's masks (Section 3.2). If the masks are wrong, the masking simulation and the area-weighted loss are corrupted.
  • domain assumption Random rectangular masking simulates real-world occlusion
    The IMVAD task is defined by this simulation (Section 3.2). No real occlusion data is used for validation, so this equivalence is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IMMoE: Incomplete Multi-View Anomaly Detection via Mixture of View Experts Fusion." pith.science (2026). https://pith.science/paper/WOHSPGSS

@misc{pith2026260719032,
  author       = {Pith},
  title        = {Pith review of: IMMoE: Incomplete Multi-View Anomaly Detection via Mixture of View Experts Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WOHSPGSS}},
  note         = {Machine review of arXiv:2607.19032}
}
read the original abstract

Existing Multi-view Anomaly Detection (MAD) methods assume that all views are completely available and model each view separately. However, in real industrial scenarios, information in the view may be missing due to faults such as occlusion, which leads to the performance degradation of existing methods due to the lack of a multi-view consistency prior. To address this, we explored a more challenging task: Incomplete Multi-View Anomaly Detection (IMVAD), in which some areas of each view were masked. We proposed a pipeline for automatically generating the IMVAD dataset and generated the \textbf{RIMAD} dataset based on the Real-IAD dataset through this pipeline. In addition, in order to effectively utilize the information of multiple views in the absence of view information, we propose \textbf{IMMoE}, which consists of two key modules: (1) Multi-View Expert Fusion (MVEF) effectively fuses multi-view information through a multi-view expert network and guides the reconstruction of a single view; (2) Local Anomaly Enhancement Encoder (LAEE) effectively prevents the model from overfitting the mask region by applying dropout to local features. Our method achieves state-of-the-art performance on both the RIMAD and Real-IAD datasets, especially on RIMAD, we have increased the pixel-level and image-level metrics by 11.8\% and 2.8\%, respectively. Our source code is available at https://github.com/HULEI7/IMMoE

Figures

Figures reproduced from arXiv: 2607.19032 by the authors.

Figure 1
Figure 1. Comparisons of our method with the previous single-view-based method on IMVAD. (a) Due to the lack of multi-view interaction, previous methods will experi￾ence significant performance degradation when partial information is missing. (b) Our method effectively integrates the useful information among various views through a multi-view expert network, thereby achieving effective information complementarity. See [PITH_… view at source ↗
Figure 2
Figure 2. The overall production process of RIMAD. We first employ Segment Anything to extract the product masks, after which we simulate view incompleteness by randomly discarding image content within selected regions. tency between different views. When a view is incomplete, the model cannot borrow useful information from other views to fill in the gaps, leading to unreli￾able detection results. (2) Overfitting of the incom… view at source ↗
Figure 3
Figure 3. The overall architecture of the IMMoE framework. Our reconstruction-based pipeline centers on two synergistic modules: the Local Anomaly Enhancement En￾coder (LAEE), which prevents overfitting to masked regions by simulating anomaly patterns through stochastic feature discarding, and the Multi-view Expert Fusion (MVEF), which adaptively aggregates cross-view complementary information via a Mixture-of-Experts mechani… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualization of anomalous region segmentation. Areas with high anomaly scores are shown in red, while areas with low anomaly scores are shown in blue [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Visualization of different mask ratio on RIMAD (default=50%). timal dropout rate is essential for balancing feature robustness with information integrity, ensuring the model focuses on learning generalizable anomaly patterns. 6 Conclusion This paper introduces IMVAD (I…
Figure 6
Figure 6. Figure 6: Visualization of anomalous region segmentation on RIMAD. Areas with high anomaly scores are shown in red, while areas with low anomaly scores are shown in blue [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 3 canonical work pages

  1. [1]

    In: Proceedings of the ieee/cvf international conference on computer vision

    Aqeel, M., Sharifi, S., Cristani, M., Setti, F.: Towards real unsupervised anomaly detection via confident meta-learning. In: Proceedings of the ieee/cvf international conference on computer vision. pp. 4858–4867 (2025)

    2025

  2. [2]

    In: Kwok, J

    Cai, W., Huang, W., Cao, Y., Huang, C., Yuan, F., Zhang, B., Wen, J.: To- wards vlm-based hybrid explainable prompt enhancement for zero-shot industrial anomaly detection. In: Kwok, J. (ed.) Proceedings of the Thirty-Fourth Interna- tional Joint Conference on Artificial Intelligence, IJCAI-25. pp. 711–719. Interna- tional Joint Conferences on Artificial I...

    work page doi:10.24963/ijcai.2025/80 2025

  3. [3]

    In: Kwok, J

    Cheng, J., Gao, C., Zhou, J., Wen, J., Dai, T., Wang, J.: Mc3d-ad: A unified geometry-aware reconstruction model for multi-category 3d anomaly detection. In: Kwok, J. (ed.) Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25. pp. 837–845. International Joint Conferences on Artificial Intelligence Organizati...

    doi:10.24963/ijcai 2025

  4. [4]

    arXiv preprint arXiv:2309.16588 (2023)

    Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need reg- isters. arXiv preprint arXiv:2309.16588 (2023)

    arXiv 2023

  5. [5]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Deng, H., Li, X.: Anomaly detection via reverse distillation from one-class embed- ding. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9737–9746 (2022)

    2022

  6. [6]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

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

    2025

  7. [7]

    Journal of Machine Learning Research 23(120), 1–39 (2022)

    Fedus,W.,Zoph,B.,Shazeer,N.:Switchtransformers:Scalingtotrillionparameter models with simple and efficient sparsity. Journal of Machine Learning Research 23(120), 1–39 (2022)

    2022

  8. [8]

    arXiv preprint arXiv:2508.06203 (2025)

    Gu, Z., Zhu, B., Zhu, G., Chen, Y., Ge, W., Tang, M., Wang, J.: Anomalymoe: Towards a language-free generalist model for unified visual anomaly detection. arXiv preprint arXiv:2508.06203 (2025)

    arXiv 2025

Show all 43 references
  1. [9]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Guo, J., Lu, S., Zhang, W., Chen, F., Li, H., Liao, H.: Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 20405–20415 (2025)

    2025

  2. [10]

    In: Proceedings of the IEEE/CVF interna- tional conference on computer vision

    Han, D., Pan, X., Han, Y., Song, S., Huang, G.: Flatten transformer: Vision trans- former using focused linear attention. In: Proceedings of the IEEE/CVF interna- tional conference on computer vision. pp. 5961–5971 (2023)

    2023

  3. [11]

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

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

    2024

  4. [12]

    In: Proceedings of the AAAI conference on artificial intelligence

    He, H., Zhang, J., Chen, H., Chen, X., Li, Z., Chen, X., Wang, Y., Wang, C., Xie, L.: A diffusion-based framework for multi-class anomaly detection. In: Proceedings of the AAAI conference on artificial intelligence. vol. 38, pp. 8472–8480 (2024)

    2024

  5. [13]

    IEEE Transactions on Multimedia (2026)

    He, H., Zhang, J., Tian, G., Wang, C., Xie, L.: Learning multi-view anomaly de- tection with efficient adaptive selection. IEEE Transactions on Multimedia (2026)

    2026

  6. [14]

    In: Kwok, J

    Hu, L., Gan, Z., Deng, L., Liang, J., Liang, L., Huang, S., Chen, T.: Replaycad: Generative diffusion replay for continual anomaly detection. In: Kwok, J. (ed.) 16 Lei Hu Proceedings of the Thirty-Fourth International Joint Conference on Artificial In- telligence, IJCAI-25. pp...

    2025 doi

  7. [15]

    In: International Conference on Pattern Recognition

    Hu, L., Huang, S.: Enhancing table structure recognition via bounding box guid- ance. In: International Conference on Pattern Recognition. pp. 209–225. Springer (2024)

    2024

  8. [16]

    In: Kwok, J

    Huang, C., Li, Q., Wen, J., Zhang, B.: Omni-dimensional state space model- driven sam for pixel-level anomaly detection. In: Kwok, J. (ed.) Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI-25. pp. 1152–1160. International Joint ...

    2025 doi

  9. [17]

    Neural computation3(1), 79–87 (1991)

    Jacobs, R.A., Jordan, M.I., Nowlan, S.J., Hinton, G.E.: Adaptive mixtures of local experts. Neural computation3(1), 79–87 (1991)

    1991

  10. [18]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4015–4026 (2023)

    2023

  11. [19]

    In: International Conference on Intelligent Computing

    Lei, T., Chen, S., Wang, B., Jiang, Z., Zou, N.: Adapted-moe: Mixture of experts with test-time adaption for anomaly detection. In: International Conference on Intelligent Computing. pp. 427–441. Springer (2025)

    2025

  12. [20]

    In: Interna- tional Conference on Pattern Recognition

    Lin, J., Chen, Z., Liang, L., Peng, W., Huang, S.: Handwriting trajectory recovery via trajectory transformer with global radical context-aware module. In: Interna- tional Conference on Pattern Recognition. pp. 182–195. Springer (2024)

    2024

  13. [21]

    Pattern Recognition179, 113679 (2026)

    Lin, J., Dai, G., Peng, W., Huang, S., Lao, Y., Zhang, H., Chen, T.: Atcmd- bench:Agentictraditionalchinesemedicinediagnosisbenchmarkforlargelanguage models via multi-agent simulation. Pattern Recognition179, 113679 (2026)

    2026

  14. [22]

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

    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)

    2024

  15. [23]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Liu, Z., Zhou, Y., Xu, Y., Wang, Z.: Simplenet: A simple network for image anomaly detection and localization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 20402–20411 (2023)

    2023

  16. [24]

    In: Proceedings of the Computer Vision and Pattern Recognition Con- ference

    Luo, W., Cao, Y., Yao, H., Zhang, X., Lou, J., Cheng, Y., Shen, W., Yu, W.: Exploring intrinsic normal prototypes within a single image for universal anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Con- ference. pp. 9974–9983 (2025)

    2025

  17. [25]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Mao, K., Lian, Y., Wang, Y., Liu, M., Zheng, N., Wei, P.: Unveiling multi-view anomaly detection: Intra-view decoupling and inter-view fusion. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 12381–12389 (2025)

    2025

  18. [27]

    In: European Conference on Computer Vision

    Meng,S.,Meng,W.,Zhou,Q.,Li,S.,Hou,W.,He,S.:Moead:Aparameter-efficient model for multi-class anomaly detection. In: European Conference on Computer Vision. pp. 345–361. Springer (2024)

    2024

  19. [28]

    In: Proceedings of the 33rd ACM International Conference on Multimedia

    Miao, B., Zhang, W., Li, J., Wu, W., Tang, S., Li, Z., Shi, H., Xiao, J., Zhuang, Y.: Robust modality-incomplete anomaly detection: A modality-instructive framework with benchmark. In: Proceedings of the 33rd ACM International Conference on Multimedia. pp. 7317–7326 (2025) IMMoE 17

    2025

  20. [29]

    International Journal of Computer Vision133(5), 2441–2462 (2025)

    Peng, W., Huang, H., Chen, T., Ke, Q., Dai, G., Huang, S.: Globally correlation- aware hard negative generation. International Journal of Computer Vision133(5), 2441–2462 (2025)

    2025

  21. [30]

    Neural Networks p

    Peng, W., Ke, Q., Liang, J., Huang, S., Chen, T.: Proxy-an loss for deep metric learning. Neural Networks p. 108254 (2025)

    2025

  22. [31]

    arXiv preprint arXiv:1904.09237 (2019)

    Reddi, S.J., Kale, S., Kumar, S.: On the convergence of adam and beyond. arXiv preprint arXiv:1904.09237 (2019)

    1904 arXiv

  23. [32]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Riquelme, C., Puigcerver, J., Nayak, B., Mariet, Z., Pinto, M., et al.: Scaling vision with sparse mixture of experts. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 34, pp. 8583–8595 (2021)

    2021

  24. [33]

    arXiv preprint arXiv:2511.18116 (2025)

    Shao, Y., Wang, L., Li, C., Chen, P., Liu, Q.: Promptmoe: Generalizable zero- shot anomaly detection via visually-guided prompt mixtures. arXiv preprint arXiv:2511.18116 (2025)

    2025

  25. [34]

    arXiv preprint arXiv:1701.06538 (2017)

    Shazeer,N.,Mirhovseini,A.,Maziarz,K.,Davis,A.,Quoc,L.,Hinton,G.,Dean,J.: Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538 (2017)

    2017 arXiv

  26. [35]

    In: Proceedings of the IEEE/CVF winter conference on appli- cations of computer vision

    Shen, Z., Zhang, M., Zhao, H., Yi, S., Li, H.: Efficient attention: Attention with linear complexities. In: Proceedings of the IEEE/CVF winter conference on appli- cations of computer vision. pp. 3531–3539 (2021)

    2021

  27. [36]

    The journal of machine learning research15(1), 1929–1958 (2014)

    Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R.: Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research15(1), 1929–1958 (2014)

    1929

  28. [37]

    Advances in neural information pro- cessing systems30(2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017)

    2017

  29. [38]

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

    Wang, C., Zhu, W., Gao, B.B., Gan, Z., Zhang, J., Gu, Z., Qian, S., Chen, M., Ma, L.: Real-iad: A real-world multi-view dataset for benchmarking versatile industrial anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2...

    2024

  30. [39]

    Advances in Neural Information Processing Systems36, 10271–10298 (2023)

    Wortsman,M.,Dettmers,T.,Zettlemoyer,L.,Morcos,A.,Farhadi,A.,Schmidt,L.: Stable and low-precision training for large-scale vision-language models. Advances in Neural Information Processing Systems36, 10271–10298 (2023)

    2023

  31. [40]

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

    You, Z., Cui, L., Shen, Y., Yang, K., Lu, X., Zheng, Y., Le, X.: A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems 35, 4571–4584 (2022)

    2022

  32. [41]

    In: ICCV

    Zavrtanik, V., Kristan, M., Skočaj, D.: Draem-a discriminatively trained recon- struction embedding for surface anomaly detection. In: ICCV. pp. 8330–8339 (2021)

    2021

  33. [42]

    arXiv preprint arXiv:2312.07495 (2023)

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

    2023 arXiv

  34. [43]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

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

    2023

  35. [44]

    In: SIGKDD

    Zhou, C., Paffenroth, R.C.: Anomaly detection with robust deep autoencoders. In: SIGKDD. pp. 665–674 (2017) 18 Lei Hu A Appendix This supplementary material consists of: –In Section A.1, we have provided more ablation experiments. –In Section A.2, we have provided more results...

    2017

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.