Pith. sign in

REVIEW 56 references

Treating an image and its reconstruction as two modalities and matching them locally finds subtle industrial defects more precisely than pixel differencing.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-30 18:54 UTC pith:UTNHNCZU

load-bearing objection Competent multi-class UAD systems paper: bidirectional windowed matching on input/reconstruction plus wavelet fusion works on the tables, but headline numbers disagree with Table I and everything rides on the frozen reconstructor.

arxiv 2607.23658 v1 pith:UTNHNCZU submitted 2026-07-26 cs.CV

XMatchAD: A Cross-Modal Matching Perspective on Reconstruction-based Anomaly Detection

Mingxiu Cai , Zhe Zhang , Gaochang Wu , Tianyou Chai This is my paper
classification cs.CV
keywords unsupervised anomaly detectionreconstruction-based methodscross-modal matchingwindowed cross-attentionfrequency-aware fusionmulti-class industrial inspectionanomaly localization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Reconstruction-based anomaly detectors often miss faint defects and smear their edges because they only compare an image to a cleaned-up version with crude similarity scores. This paper claims that if you instead treat the original and the reconstruction as two complementary “modalities,” run local windowed cross-attention between them, and fuse their high-frequency detail bands, you obtain sharper anomaly maps and higher multi-class detection scores. The method, XMatchAD, stacks self-attention inside each branch, bidirectional windowed cross-attention between branches, and a wavelet-guided fusion head that emphasizes edges and textures. On the standard industrial benchmarks MVTec-AD, VisA and MPDD it reports state-of-the-art image- and pixel-level numbers under the unified multi-class setting. A sympathetic reader cares because many factory inspection pipelines already rely on reconstruction; a better matching layer could be dropped on top without new labels.

Core claim

Reconstruction-based unsupervised anomaly detection can be reformulated as pseudo cross-modal matching: once a diffusion or lightweight reconstructor supplies an anomaly-free counterpart, local bidirectional windowed cross-attention between DINOv2 features of the two images, followed by adaptive high-frequency fusion, produces anomaly maps that are more sensitive to subtle defects and have sharper boundaries than cosine-similarity or global differencing baselines, yielding superior multi-class detection and localization on MVTec-AD, VisA and MPDD.

What carries the argument

Attention-guided cross-modal matching (window-based multi-head self- and cross-attention, W-MSA/W-MCA) plus adaptive frequency-aware fusion: the former locally aligns and mutually refines the two modality streams; the latter couples high-frequency wavelet sub-bands across scales to sharpen anomaly boundaries.

Load-bearing premise

The frozen reconstructor must produce sufficiently clean, anomaly-free images so that residual mismatches after feature encoding are true defects rather than reconstruction artifacts.

What would settle it

Replace the reconstructor with one that systematically leaves residual defects or invents structure on known anomalous images; if image- and pixel-AUROC then collapse toward the plain cosine-similarity baseline, the central matching claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Factory inspection systems that already run a reconstructor can gain sharper localization by swapping the final differencing step for local cross-modal matching.
  • High-frequency wavelet guidance becomes a reusable recipe for boundary-sensitive anomaly heads on other reconstruction pipelines.
  • Multi-class unified models can reduce false negatives on small or low-contrast defects without class-specific training.
  • Inference cost is dominated by reconstruction speed; faster reconstructors immediately shorten the end-to-end latency of the full pipeline.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same local-matching-plus-frequency-fusion block could be attached to embedding-only teachers that never reconstruct, testing whether the reconstructor is essential or merely convenient.
  • If window size is made adaptive to estimated defect scale, the method might close remaining gaps on extremely small or elongated anomalies.
  • Bidirectional matching scores themselves could serve as an unsupervised signal for active sampling of hard normal variants during training.

Editorial analysis

A structured set of objections, weighed in public.

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

Circularity Check

0 steps flagged

No significant circularity: standard reconstruction-based UAD architecture evaluated on external benchmarks; metrics are not forced by construction.

full rationale

XMatchAD reframes reconstruction-based unsupervised anomaly detection as pseudo cross-modal matching between an input and a frozen reconstructor’s output (GLAD or Dinomaly), then trains windowed self/cross-attention and frequency-aware fusion with BCE against synthetic anomaly masks and reports I-/P-AUROC/AP/F1 on held-out MVTec-AD, VisA, and MPDD splits against external SOTA. The headline numbers are empirical test-set outcomes, not algebraic restatements of fitted constants. Loss head weights (α, β, γ) and architectural choices (W-MSA/W-MCA, high-frequency DWT coupling) are design hyperparameters ablated on the same external splits; they do not define the reported AUROC by construction. Self-citations (e.g., CostFilter-AD) appear only as related work and are not load-bearing uniqueness or uniqueness-theorem premises. The reconstructor quality assumption is a modeling premise, not a circular derivation. Derivation chain is self-contained engineering plus external evaluation; score 0.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

Empirical CV method paper. Central performance claims rest on standard UAD protocol assumptions, frozen foundation features, external reconstructors, and several hand-chosen hyperparameters rather than on novel physical entities. No formal derivation; inductive bias is that local inter-view mismatch after reconstruction equals anomaly.

free parameters (5)
  • loss weights α, β, γ = 0.8, 0.1, 0.1
    Balance Lf, Lt→r, Lr→t in Eq. (14); set empirically to 0.8, 0.1, 0.1 without sweep reported as primary evidence.
  • top-K pixels K for image score = 250
    Image-level score averages top-K values of Ms; K chosen empirically.
  • architecture scales L, C′, H′, W′ = L=3, C′=48, H′=W′=28
    Matching depth and feature map shape set by hand for the 256 setting.
  • learning rate = 4e-3
    Optimizer step size chosen empirically.
  • GLAD denoising steps = 25
    Reconstruction quality/cost knob fixed at 25 steps for the main setting.
axioms (5)
  • domain assumption Models trained only on normal data, with synthetic anomaly supervision for the segmentation heads, generalize to real industrial defects under multi-class evaluation.
    Standard UAD protocol invoked throughout §I and §IV-A; self-supervised strategy “as commonly used in UAD methods.”
  • domain assumption A pretrained DINOv2 encoder provides sufficiently discriminative frozen features for both input and reconstruction branches.
    §III-A uses 12th-layer (or Dinomaly fused) DINOv2 features without task-specific backbone training.
  • domain assumption Anomalies are spatially localized and continuous, so windowed attention is an appropriate inductive bias versus global attention.
    Stated in §I and §III-B to justify shift-window W-MSA/W-MCA over vanilla global matching.
  • ad hoc to paper High-frequency wavelet subbands carry the boundary/texture cues needed for sharp anomaly delineation; low-frequency content is less useful or harmful.
    §III-C and Table VI motivate dropping low-frequency subbands after DWT; supported empirically here but not a universal law.
  • standard math Softmax attention and residual MLP updates are valid differentiable matching operators replacing cosine similarity (Eq. 1 → Eqs. 4–6).
    Standard transformer algebra used as the learnable matching mechanism.
invented entities (2)
  • Pseudo cross-modal matching formulation (input vs reconstruction as modalities) no independent evidence
    purpose: Reframe residual AD as bidirectional inter-view matching to justify W-MCA and dual-branch refinement.
    Same-sensor pair labeled as two modalities; useful design story but not an independently measured modality or physical object.
  • Frequency-aware Cross-modal Fusion (FCF) with product-coupled high-frequency gates no independent evidence
    purpose: Produce adaptive weights from multi-scale high-frequency DWT features to sharpen Mf.
    Architectural module introduced in §III-C; evidence is ablation gains on VisA, not external measurement.

pith-pipeline@v1.2.0-grok45-kimik3 · 25893 in / 3628 out tokens · 71388 ms · 2026-07-30T18:54:50.258233+00:00 · methodology

0 comments
read the original abstract

The remarkable success of reconstruction-based methods in Unsupervised Anomaly Detection (UAD) lies in their ability to identify and localize anomalies by modeling discrepancies between input images and their reconstructed counterparts. However, these approaches often struggle to capture subtle anomalies and tend to produce blurred anomaly boundaries, which significantly limits their effectiveness, particularly in complex multi-class scenarios. To address these issues, we present XMatchAD, a novel UAD framework that reinterprets the task from a pseudo cross-modal matching perspective. Specifically, the input and reconstructed images are treated as two complementary modalities and their matching relationships are precisely exploited for anomaly detection. First, a pre-trained feature extractor is employed to encode discriminative representations. Second, an attention-guided cross-modal matching mechanism is introduced to match local inter-modal anomaly-related patterns while mutually refining the features. This enhances the sensitivity to anomalies with diverse shapes and subtle deviations and significantly improves the precision of anomaly detection and localization. Third, we design an adaptive frequency-aware fusion module that further delineates sharp anomaly boundaries through the coupling of high-frequency components from cross-modal multi-scale representations. Comprehensive evaluations on MVTec-AD, VisA, and MPDD benchmarks demonstrate that our method consistently achieves superior performance, outperforming state-of-the-art methods in multi-class anomaly detection and localization. The code will be released at https://github.com/Mingxiu-Cai/XMatchAD.

Figures

Figures reproduced from arXiv: 2607.23658 by Gaochang Wu, Mingxiu Cai, Tianyou Chai, Zhe Zhang.

Figure 1
Figure 1. Figure 1: Visualization of the predicted anomaly heatmaps from [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed XMatchAD. We reformulate UAD as a pseudo cross-modal matching task, where the input [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Design of the adaptive frequency-aware fusion. (a) Frequency-aware Cross-modal Fusion (FCF). (b) Wavelet-based [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Pixel-level anomaly score distributions for each cat [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Anomaly localization visualization on MVTec-AD and VisA datasets. From left to right are input images, ground truth, [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Anomaly localization visualization on MPDD. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Matching correlation visualization on MVTec-AD. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 1 linked inside Pith

  1. [1]

    Aadc-net: A multimodal deep learning framework for automatic anomaly detection in real-time surveillance,

    D. Tri Phan, V . Hoang Minh Doan, J. Choi, B. Lee, and J. Oh, “Aadc-net: A multimodal deep learning framework for automatic anomaly detection in real-time surveillance,”IEEE Transactions on Instrumentation and Measurement, vol. 74, pp. 1–13, 2025

    2025

  2. [3]

    Anomaly detection for medical images using heterogeneous auto-encoder,

    S. Lu, W. Zhang, H. Zhao, H. Liu, N. Wang, and H. Li, “Anomaly detection for medical images using heterogeneous auto-encoder,”IEEE Transactions on Image Processing, vol. 33, pp. 2770–2782, 2024

    2024

  3. [4]

    Enhancing unsupervised anomaly detection with score-guided network,

    Z. Huang, B. Zhang, G. Hu, L. Li, Y . Xu, and Y . Jin, “Enhancing unsupervised anomaly detection with score-guided network,”IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 10, pp. 14 754–14 769, 2023

    2023

  4. [5]

    Exploiting structural consistency of chest anatomy for unsupervised anomaly detection in radiography images,

    T. Xiang, Y . Zhang, Y . Lu, A. Yuille, C. Zhang, W. Cai, and Z. Zhou, “Exploiting structural consistency of chest anatomy for unsupervised anomaly detection in radiography images,”IEEE Transactions on Pat- tern Analysis and Machine Intelligence, vol. 46, no. 9, pp. 6070–6081, 2024

    2024

  5. [6]

    Unistad: An unified triple-tower student–teacher model for multi-class anomaly detection and localization,

    H. Liu and J. Sun, “Unistad: An unified triple-tower student–teacher model for multi-class anomaly detection and localization,”IEEE Trans- actions on Circuits and Systems for Video Technology, vol. 35, no. 4, pp. 3196–3208, 2025

    2025

  6. [7]

    Boosting global- local feature matching via anomaly synthesis for multi-class point cloud anomaly detection,

    Y . Cheng, Y . Cao, D. Wang, W. Shen, and W. Li, “Boosting global- local feature matching via anomaly synthesis for multi-class point cloud anomaly detection,”IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 12 560–12 571, 2025

    2025

  7. [8]

    Costfilter- ad: Enhancing anomaly detection through matching cost filtering,

    Z. Zhang, M. Cai, H. Wang, G. Wu, T. Chai, and X. Zhu, “Costfilter- ad: Enhancing anomaly detection through matching cost filtering,”arXiv preprint arXiv:2505.01476, 2025

    arXiv 2025

  8. [9]

    Revisiting reverse distillation for anomaly detection,

    T. D. Tien, A. T. Nguyen, N. H. Tran, T. D. Huy, S. Duong, C. D. T. Nguyen, and S. Q. Truong, “Revisiting reverse distillation for anomaly detection,” inCVPR, 2023, pp. 24 511–24 520

    2023

  9. [10]

    A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization,

    Q. Chen, H. Luo, C. Lv, and Z. Zhang, “A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization,” inECCV, 2024, pp. 37–54

    2024

  10. [11]

    Unsupervised anomaly segmentation via deep feature reconstruction,

    Y . Shi, J. Yang, and Z. Qi, “Unsupervised anomaly segmentation via deep feature reconstruction,”Neurocomputing, vol. 424, pp. 9–22, 2021

    2021

  11. [12]

    Towards visually explaining variational autoencoders,

    W. Liu, R. Li, M. Zheng, S. Karanam, Z. Wu, B. Bhanu, R. J. Radke, and O. Camps, “Towards visually explaining variational autoencoders,” inCVPR, 2020, pp. 8642–8651

    2020

  12. [13]

    Divide- and-assemble: Learning block-wise memory for unsupervised anomaly detection,

    J. Hou, Y . Zhang, Q. Zhong, D. Xie, S. Pu, and H. Zhou, “Divide- and-assemble: Learning block-wise memory for unsupervised anomaly detection,” inICCV, 2021, pp. 8791–8800

    2021

  13. [14]

    Memorizing normality to detect anomaly: Memory- augmented deep autoencoder for unsupervised anomaly detection,

    D. Gong, L. Liu, V . Le, B. Saha, M. R. Mansour, S. Venkatesh, and A. v. d. Hengel, “Memorizing normality to detect anomaly: Memory- augmented deep autoencoder for unsupervised anomaly detection,” in ICCV, 2019, pp. 1705–1714

    2019

  14. [15]

    Omni-frequency channel-selection representations for unsupervised anomaly detection,

    Y . Liang, J. Zhang, S. Zhao, R. Wu, Y . Liu, and S. Pan, “Omni-frequency channel-selection representations for unsupervised anomaly detection,” IEEE Transactions on Image Processing, vol. 32, pp. 4327–4340, 2023

    2023

  15. [16]

    Reconstruction by inpainting for visual anomaly detection,

    V . Zavrtanik, M. Kristan, and D. Sko ˇcaj, “Reconstruction by inpainting for visual anomaly detection,”Pattern Recognition, vol. 112, p. 107706, 2021

    2021

  16. [17]

    Unsupervised surface anomaly detection with diffusion probabilistic model,

    X. Zhang, N. Li, J. Li, T. Dai, Y . Jiang, and S.-T. Xia, “Unsupervised surface anomaly detection with diffusion probabilistic model,” inICCV, 2023, pp. 6782–6791

    2023

  17. [18]

    Glad: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection,

    H. Yao, M. Liu, Z. Yin, Z. Yan, X. Hong, and W. Zuo, “Glad: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection,” inECCV, 2024, pp. 1–17

    2024

  18. [19]

    A diffusion-based framework for multi-class anomaly detection,

    H. He, J. Zhang, H. Chen, X. Chen, Z. Li, X. Chen, Y . Wang, C. Wang, and L. Xie, “A diffusion-based framework for multi-class anomaly detection,” inAAAI, vol. 38, no. 8, 2024, pp. 8472–8480

    2024

  19. [20]

    Adtr: Anomaly detection transformer with feature reconstruction,

    Z. You, K. Yang, W. Luo, L. Cui, Y . Zheng, and X. Le, “Adtr: Anomaly detection transformer with feature reconstruction,” inICONIP, 2022, pp. 298–310

    2022

  20. [21]

    Draem-a discriminatively trained reconstruction embedding for surface anomaly detection,

    V . Zavrtanik, M. Kristan, and D. Sko ˇcaj, “Draem-a discriminatively trained reconstruction embedding for surface anomaly detection,” in ICCV, 2021, pp. 8330–8339

    2021

  21. [22]

    Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection,

    H. Zhang, Z. Wang, D. Zeng, Z. Wu, and Y .-G. Jiang, “Diffusionad: Norm-guided one-step denoising diffusion for anomaly detection,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 8, pp. 7140–7152, 2025

    2025

  22. [23]

    Dsr–a dual subspace re- projection network for surface anomaly detection,

    V . Zavrtanik, M. Kristan, and D. Sko ˇcaj, “Dsr–a dual subspace re- projection network for surface anomaly detection,” inECCV, 2022, pp. 539–554

    2022

  23. [24]

    Unsu- pervised anomaly detection via masked diffusion posterior sampling,

    D. Wu, S. Fan, X. Zhou, L. Yu, Y . Deng, J. Zou, and B. Lin, “Unsu- pervised anomaly detection via masked diffusion posterior sampling,” arXiv preprint arXiv:2404.17900, 2024

    Pith/arXiv arXiv 2024

  24. [25]

    Anomaly detection with conditioned denoising diffusion models,

    A. Mousakhan, T. Brox, and J. Tayyub, “Anomaly detection with conditioned denoising diffusion models,” inDAGM German Conference on Pattern Recognition, 2024, pp. 181–195

    2024

  25. [26]

    A unified model for multi-class anomaly detection,

    Z. You, L. Cui, Y . Shen, K. Yang, X. Lu, Y . Zheng, and X. Le, “A unified model for multi-class anomaly detection,”NeurIPS, vol. 35, pp. 4571–4584, 2022

    2022

  26. [27]

    Destseg: Segmentation guided denoising student-teacher for anomaly detection,

    X. Zhang, S. Li, X. Li, P. Huang, J. Shan, and T. Chen, “Destseg: Segmentation guided denoising student-teacher for anomaly detection,” inCVPR, 2023, pp. 3914–3923

    2023

  27. [28]

    Just noticeable learning for unsupervised anomaly localization and detection,

    Y . Zhao, “Just noticeable learning for unsupervised anomaly localization and detection,” inICME, 2022, pp. 01–06

    2022

  28. [30]

    Learning semantic context from normal samples for unsupervised anomaly detection,

    X. Yan, H. Zhang, X. Xu, X. Hu, and P.-A. Heng, “Learning semantic context from normal samples for unsupervised anomaly detection,” in AAAI, vol. 35, no. 4, 2021, pp. 3110–3118

    2021

  29. [31]

    A survey on multimodal large language models for autonomous driving,

    C. Cui, Y . Ma, X. Cao, W. Ye, Y . Zhou, K. Liang, J. Chen, J. Lu, Z. Yang, K.-D. Liaoet al., “A survey on multimodal large language models for autonomous driving,” inWACV, 2024, pp. 958–979

    2024

  30. [32]

    Spatio-contextual deep network-based multimodal pedestrian detection for autonomous driving,

    K. Dasgupta, A. Das, S. Das, U. Bhattacharya, and S. Yogamani, “Spatio-contextual deep network-based multimodal pedestrian detection for autonomous driving,”IEEE transactions on intelligent transportation systems, vol. 23, no. 9, pp. 15 940–15 950, 2022

    2022

  31. [33]

    Unified domain adaptive semantic segmentation,

    Z. Zhang, G. Wu, J. Zhang, X. Zhu, D. Tao, and T. Chai, “Unified domain adaptive semantic segmentation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 47, no. 8, pp. 6731–6748, 2025

    2025

  32. [34]

    Multi-granularity con- trastive cross-modal collaborative generation for end-to-end long-term video question answering,

    T. Yu, K. Fu, J. Zhang, Q. Huang, and J. Yu, “Multi-granularity con- trastive cross-modal collaborative generation for end-to-end long-term video question answering,”IEEE Transactions on Image Processing, vol. 33, pp. 3115–3129, 2024

    2024

  33. [35]

    History aware multimodal transformer for vision-and-language navigation,

    S. Chen, P.-L. Guhur, C. Schmid, and I. Laptev, “History aware multimodal transformer for vision-and-language navigation,”NeurIPS, vol. 34, pp. 5834–5847, 2021

    2021

  34. [36]

    Cross-modal map learning for vision and language navigation,

    G. Georgakis, K. Schmeckpeper, K. Wanchoo, S. Dan, E. Miltsakaki, D. Roth, and K. Daniilidis, “Cross-modal map learning for vision and language navigation,” inCVPR, 2022, pp. 15 460–15 470. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 13

    2022

  35. [37]

    An overview of deep learning methods for multimodal medical data mining,

    F. Behrad and M. S. Abadeh, “An overview of deep learning methods for multimodal medical data mining,”Expert Systems with Applications, vol. 200, p. 117006, 2022

    2022

  36. [38]

    Plug-and-play regula- tors for image-text matching,

    H. Diao, Y . Zhang, W. Liu, X. Ruan, and H. Lu, “Plug-and-play regula- tors for image-text matching,”IEEE Transactions on Image Processing, vol. 32, pp. 2322–2334, 2023

    2023

  37. [39]

    Rfnet: Region-aware fusion network for incomplete multi-modal brain tumor segmentation,

    Y . Ding, X. Yu, and Y . Yang, “Rfnet: Region-aware fusion network for incomplete multi-modal brain tumor segmentation,” inICCV, 2021, pp. 3975–3984

    2021

  38. [40]

    Equivariant multi-modality image fusion,

    Z. Zhao, H. Bai, J. Zhang, Y . Zhang, K. Zhang, S. Xu, D. Chen, R. Timofte, and L. Van Gool, “Equivariant multi-modality image fusion,” inCVPR, 2024, pp. 25 912–25 921

    2024

  39. [41]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMICCAI, 2015, pp. 234–241

    2015

  40. [42]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” inCVPR, 2022, pp. 5728–5739

    2022

  41. [43]

    Gafusion: Adaptive fusing lidar and camera with multiple guidance for 3d object detection,

    X. Li, B. Fan, J. Tian, and H. Fan, “Gafusion: Adaptive fusing lidar and camera with multiple guidance for 3d object detection,” inCVPR, 2024, pp. 21 209–21 218

    2024

  42. [44]

    Event-assisted low-light video object segmentation,

    H. Li, J. Wang, J. Yuan, Y . Li, W. Weng, Y . Peng, Y . Zhang, Z. Xiong, and X. Sun, “Event-assisted low-light video object segmentation,” in CVPR, 2024, pp. 3250–3259

    2024

  43. [45]

    Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval,

    D. Jiang and M. Ye, “Cross-modal implicit relation reasoning and aligning for text-to-image person retrieval,” inCVPR, 2023, pp. 2787– 2797

    2023

  44. [46]

    See more and know more: Zero-shot point cloud segmentation via multi-modal visual data,

    Y . Lu, Q. Jiang, R. Chen, Y . Hou, X. Zhu, and Y . Ma, “See more and know more: Zero-shot point cloud segmentation via multi-modal visual data,” inICCV, 2023, pp. 21 674–21 684

    2023

  45. [47]

    Align before fuse: Vision and language representation learning with momentum distillation,

    J. Li, R. Selvaraju, A. Gotmare, S. Joty, C. Xiong, and S. C. H. Hoi, “Align before fuse: Vision and language representation learning with momentum distillation,”NeurIPS, vol. 34, pp. 9694–9705, 2021

    2021

  46. [48]

    Emerging properties in self-supervised vision transformers,

    M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging properties in self-supervised vision transformers,” inICCV, 2021, pp. 9650–9660

    2021

  47. [49]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” inICCV, 2021, pp. 10 012–10 022

    2021

  48. [50]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”NeurIPS, vol. 30, 2017

    2017

  49. [51]

    Wavelet based image fusion techniques—an introduction, review and comparison,

    K. Amolins, Y . Zhang, and P. Dare, “Wavelet based image fusion techniques—an introduction, review and comparison,”ISPRS Journal of photogrammetry and Remote Sensing, vol. 62, no. 4, pp. 249–263, 2007

    2007

  50. [52]

    Xnet: Wavelet- based low and high frequency fusion networks for fully-and semi- supervised semantic segmentation of biomedical images,

    Y . Zhou, J. Huang, C. Wang, L. Song, and G. Yang, “Xnet: Wavelet- based low and high frequency fusion networks for fully-and semi- supervised semantic segmentation of biomedical images,” inICCV, 2023, pp. 21 085–21 096

    2023

  51. [53]

    Simplenet: A simple network for image anomaly detection and localization,

    Z. Liu, Y . Zhou, Y . Xu, and Z. Wang, “Simplenet: A simple network for image anomaly detection and localization,” inCVPR, 2023, pp. 20 402– 20 411

    2023

  52. [54]

    Exploring plain vit features for multi-class unsupervised visual anomaly detection,

    J. Zhang, X. Chen, Y . Wang, C. Wang, Y . Liu, X. Li, M.-H. Yang, and D. Tao, “Exploring plain vit features for multi-class unsupervised visual anomaly detection,”Computer Vision and Image Understanding, vol. 253, p. 104308, 2025

    2025

  53. [55]

    Anomalydino: Boosting patch-based few-shot anomaly detection with dinov2,

    S. Damm, M. Laszkiewicz, J. Lederer, and A. Fischer, “Anomalydino: Boosting patch-based few-shot anomaly detection with dinov2,” in WACV. IEEE, 2025, pp. 1319–1329

    2025

  54. [56]

    Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detec- tion,

    J. Guo, S. Lu, W. Zhang, F. Chen, H. Li, and H. Liao, “Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detec- tion,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 20 405–20 415

    2025

  55. [57]

    On estimation of a probability density function and mode,

    E. Parzen, “On estimation of a probability density function and mode,” The annals of mathematical statistics, vol. 33, no. 3, pp. 1065–1076, 1962

    1962

  56. [58]

    Grad-cam: Visual explanations from deep networks via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” inICCV, 2017, pp. 618–626. Mingxiu Caireceived the MS degree from the School of Computer Science, Nanjing Audit Uni- versity, China, in 2024. She is currently working toward a Ph.D. degree in the S...

    2017