Pith. sign in

REVIEW 5 major objections 5 minor 59 references

PADFormer: Pose-agnostic Anomaly Detection from Sparse View Images

T0 review · 5 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Masked reconstruction from a few unaligned views can spot defects at any viewpoint without 3D reconstruction.

desk verdict Genuinely new sparse-view PAD method with sound core ideas, but the headline state-of-the-art claim is overstated by weak baselines and missing error bars. read the letter →

arxiv 2608.04210 v1 pith:OCESDKSU submitted 2026-08-04 cs.CV

Ruiqi Wang , Yiming Qian , Fenggen Yu , Yuxuan Lu , Dakuo Wang , Hao Zhang , Jing Huang This is my paper · ORCID
classification cs.CV
keywords pose-agnosticanomalydetectionfew-shotmaskedimagemodelingvisiontransformercross-viewreconstructionspatialnetworkMAD-SIMMVTec-AD
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

PADFormer sets out to show that pose-agnostic anomaly detection—finding defects in objects photographed from arbitrary, unknown viewpoints—does not require building a 3D model from hundreds of posed reference images. The paper's claim is that a Vision Transformer trained purely on anomaly-free images can reconstruct an anomaly-free version of a query image directly in 2D, preserving the query's viewpoint, by masking part of the query and filling it from a few unaligned reference views. On the MAD-SIM benchmark this raises image-level AUROC from the best published baseline's 57.9 to 78.8 with two references, from 60.3 to 81.3 with four, and from 58.2 to 85.6 with ten, while running inference in under a second and without per-object training. A sympathetic reader would care because the assumption of aligned viewpoints or dense multi-view capture is the main obstacle to deploying anomaly detection in real inspection lines.

What carries the argument

Three components work together. First, cross-view masked reconstruction: the query is tokenised into patches, 40% are randomly masked, the visible patches pass through a ViT encoder, all reference images are encoded without masking, and a cross-attention decoder must recover the masked patches from the references plus learnable tokens. Second, dynamic patch selection with a Spatial Transformer Network (STN): for each reference image the STN predicts a 6-parameter affine transform that warps the reference feature grid toward the query grid, then for every query position the model keeps the top-10 most similar aligned reference patches as cross-attention keys and values. Third, anomaly-agnostic tokens: 32 global tokens and one position-indexed local token per patch carry normal-appearance priors that do not depend on any specific defect type. During inference, 15 different random masks are applied, the per-pass reconstruction errors are averaged in CIELAB space, and a variance-based mask suppresses consistent reconstruction artifacts so genuine anomalies stand out.

What would settle it

Hold out test queries on MAD-SIM whose nearest reference views are at least 60 degrees apart in azimuth from the query; if PADFormer's image-AUROC collapses toward the reported 76.2 worst-case value, the affine alignment ceiling is confirmed. A second test: feed the same reference images to PADFormer but give the STN ground-truth relative poses instead of predicted warps; if performance does not rise above 81.3, the reconstruction machinery rather than alignment is the binding constraint.

Watch

Extended reading notes

Core claim

PADFormer's central discovery is that masked reconstruction trained exclusively on normal data generalises to novel, previously unseen viewpoints: when a query image is randomly masked and a decoder is asked to recover the missing patches using cross-attention to a sparse set of anomaly-free reference images, the network does not reproduce an anomaly—it replaces it with plausible normal content—yet it preserves the exact pose of the query. The paper claims this makes anomaly detection a simple comparison between the query and its reconstructed anomaly-free counterpart, with pose robustness emerging without any explicit pose estimation, 3D reconstruction, or test-time alignment optimization. The paper further claims the mechanism is sample-efficient: with only 2, 4, or 10 reference images it outperforms 3D-reconstruction baselines by a large margin on the PAD benchmark, and a single model trained jointly across all object categories transfers to classic few-shot anomaly detection on MVTec-AD and VisA at a competitive level.

Load-bearing premise

The load-bearing premise is that the STN's six-parameter affine warp, computed from patch-level features, can align sparse reference views to the query view closely enough for top-k patch retrieval under large out-of-plane rotations; the paper's own worst-case reference experiment, where three of four references come from opposite sides, drops 4-shot image-AUROC from 81.3 to 76.2, so alignment quality is the fragile link.

Editorial extensions

If this is right

  • Sparse-view PAD becomes practical: a handful of unposed reference photos per object replaces the 200+ posed views and per-object 3D fitting that earlier 3D-reconstruction methods require.
  • Inference cost drops to about 0.7–0.8 seconds per query at 4- and 10-shot, 4–50× faster than the 3D baselines, and no test-time pose optimisation is needed.
  • A single model trained jointly across all object categories reaches the top image-level AUROC in the all-shot PAD setting, suggesting the learned normal-prior representation is not tied to one object class.
  • The same architecture, without modification, reaches competitive performance on MVTec-AD and VisA few-shot anomaly detection, where viewpoints are aligned, so the method covers both the pose-agnostic and the aligned-viewpoint settings.

Reading between the lines

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

  • Editorial inference: the whole approach can be read as a general recipe—train a masked reconstructor on normal data of any object category and it will 'normal-ise' anomalies in novel views; the same recipe might extend to non-rigid objects such as fabric or food, provided the STN alignment is replaced or augmented with a richer warp model.
  • Editorial inference: the failure mode under opposite-side references suggests the affine alignment ceiling, not the reconstruction idea itself, is the current bottleneck; adding a lightweight pose estimate or a coarse 3D prior at the feature level could move the worst-case 76.2 image-AUROC closer to the best-case 84.2.
  • Editorial inference: the variance-based consistency mask separates persistent reconstruction bias from stochastic anomaly response by a hand-set median threshold; a testable extension is to learn this mask, which could improve pixel-level localisation on reflective or cluttered surfaces, as the paper's MAD-Real results hint.
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

5 major / 5 minor

Summary. The paper proposes PADFormer, a ViT-based method for pose-agnostic anomaly detection (PAD) that operates from a sparse set of anomaly-free reference images (2, 4, or 10) without 3D reconstruction or test-time pose optimization. The method reconstructs an anomaly-free version of a query image by masked cross-view reconstruction, using a Spatial Transformer Network (STN) to align reference patch features, dynamic patch selection to retrieve the most relevant reference patches, and learnable anomaly-agnostic tokens as normal priors. At inference, multiple reconstruction passes with different masks are averaged, and an anomaly score is computed from a CIELAB error map with a variance-based consistency mask. The paper reports state-of-the-art results on the MAD-SIM and PIAD benchmarks in the sparse-view setting, competitive results on FSAD benchmarks (MVTec-AD and VisA), and a full set of ablations on the proposed modules, the error metric, training strategy, and hyperparameters.

Significance. If the sparse-view PAD results are reliable, the paper would make a useful practical contribution: it removes the dense-view, pose-annotated, per-object training requirement of existing PAD methods and replaces 3D reconstruction with image-space masked reconstruction. The architecture is clearly described, and the supplementary material includes useful ablations (k-sensitivity, reference-selection analysis, two-stage training, consistency-mask effect). However, the claimed state-of-the-art margin is not established: the headline comparison is against dense-view 3D reconstruction baselines run outside their operating regime, while the paper's own architecture-matched baseline shows only a ~4.5-point gain. Missing error bars, an unanalyzed consistency-mask threshold, and a large MAD-Real discrepancy further weaken the quantitative claims.

major comments (5)
  1. [Sec. 4.3, Table 1] The headline comparison is not on equal footing. OmniAD, SplatPose, and PIAD are dense-view 3D reconstruction methods designed for hundreds of pose-annotated references and per-object training; feeding them 2/4/10 images forces them to run far outside their intended setting, so their low scores are expected rather than informative. The only architecture-matched sparse-view baseline appears in Table 3 (the CroCo-style cross-attention decoder trained on anomaly-free data), which reaches 76.8 image-AUROC on 4-shot MAD-SIM versus PADFormer's 81.3. The actual algorithmic margin is therefore ~4.5 points, not the >20-point gap over the best published baseline highlighted in the abstract. Please include a properly tuned sparse-view baseline that uses the same multi-pass LAB error metric and consistency mask, and frame the state-of-the-art claim accordingly.
  2. [Tables 1-4, Sec. 3.3] No uncertainty is reported for any experimental result. The pipeline has multiple stochastic components: random reference selection (Sec. 4.1 and Sec. 3.3), random masking patterns (Sec. 3.3), and multiple inference passes (I=15). The claimed margins, e.g., 81.3 vs 76.8 in Table 3, may be within run-to-run variability. Please report mean and standard deviation over at least three independent runs (with different random seeds/reference draws) for the main PAD and FSAD tables, as well as for the ablation tables.
  3. [Sec. 3.3, Eq. (6)] The consistency mask is a load-bearing component of the anomaly score, and the paper reports it contributes +2.4% image-AUROC (Sec. 4.4, last sentence). However, the threshold P50 is a free parameter with no sensitivity analysis, and the assertion that high variance corresponds to genuine anomalies while low variance corresponds to rare-but-normal patterns or illumination drift is only a heuristic. Please add an ablation over the variance threshold (e.g., P25, P50, P75) and, ideally, a controlled experiment showing that consistent reconstruction artifacts are suppressed while true anomalies are retained.
  4. [Sec. 4.3, Table 2] The claim of "comparable performance" on classic FSAD tasks is not supported by the numbers. On MVTec-AD and VisA, PADFormer is consistently 2-4 points below UniVAD across all shot settings (e.g., 94.7 vs 98.6 in 4-shot MVTec image-AUROC) and also below WinCLIP in most settings. The paper argues a task mismatch, but the word "comparable" is an overstatement; please soften the claim or provide a statistical comparison (e.g., paired tests over categories).
  5. [Sec. 4.1 footnote; Supplementary Table (MAD-Real)] The paper's footnote says MAD-Real was not used because OmniAD recommends MAD-Sim, but the supplementary reports a MAD-Real evaluation: PADFormer (4-shot, with segmentation) reaches 80.4 image-AUROC versus OmniAD (50-shot) at 81.0. This demonstrates that the real-world advantage over a dense-view baseline is much smaller than the synthetic MAD-SIM results suggest. The main text should discuss this discrepancy explicitly and justify why the synthetic benchmark is the appropriate basis for the state-of-the-art claim.
minor comments (5)
  1. [Sec. 4.3] Typo: "wheras" should be "whereas".
  2. [Sec. 4.3, Table 1] The text says "over 20%" and "8%" but the numbers are absolute percentage points (e.g., 78.8 vs 57.9 is 20.9 points, not 20.9% relative). Please use "points" for clarity.
  3. [Supplementary Sec. A] The description of the decoder projections and unmasking is repeated almost verbatim in two consecutive paragraphs; this appears to be a copy-paste error and should be cleaned up.
  4. [Throughout] Please use a consistent spelling for the dataset ("MAD-SIM" vs "MAD-Sim") and for the author names in the references (e.g., "Skočaj" appears with an accent; keep the original diacritics).
  5. [Sec. 3.3, inference] The choice of mask ratio r=0.4 and I=15 is justified by a coverage probability calculation, which is good; the supplementary ablation (Table 7) confirms the choice. Please consider moving this ablation into the main text because it is central to the method's inference cost.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: PADFormer is trained on anomaly-free data, evaluated on held-out queries, and does not reduce any prediction to a fitted input or self-citation by construction.

full rationale

PADFormer's central mechanism is cross-view masked reconstruction trained with the loss in Eq. 4 on anomaly-free query/reference pairs only. At inference, Eq. 5 computes per-pass CIELAB errors between the query and its reconstructions, and Eq. 6 applies a variance-based consistency mask. The pipeline never fits test labels, and the reported AUROC numbers are measured on held-out abnormal and normal queries rather than being recovered from the training objective. The hyperparameters (mask ratio 0.4, top-k=10, 15 inference passes, P50 variance threshold) are fixed and ablated on MAD-SIM, but none of them is equivalent to the final anomaly score by construction. No load-bearing self-citation is present: the cited dataset [55], CroCo [46], MAE [15], STN [20], and the compared baselines are all external works, and the paper does not invoke any uniqueness theorem or prior-work premise to force its design. The Table 3 ablation starts from a CroCo-style baseline (76.8 image-AUROC) and adds STN, DPS, and AAT modules, which is an honest decomposition rather than a circular restatement of the final result. The supplementary worst-case reference analysis (76.2 vs 81.3 for random references) and the MAD-Real results (80.4 with segmentation vs 81.0 for OmniAD with 50 shots) are reported as limitations and do not conceal a definitional equivalence. Concerns about baselines being evaluated outside their intended dense-view regime are legitimate benchmarking/correctness risks, but they are not circularity: the method's predictions are not constructed from the baseline scores or from the test labels. Therefore the derivation chain is self-contained and no circular step can be exhibited.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The method's central claims rest on several hand-chosen hyperparameters (mask ratio, number of inference passes, k, w, sigma, variance threshold) and on the assumption that synthetic/semi-synthetic benchmarks reflect real deployment conditions. The only invented entity, the anomaly-agnostic tokens, is a learned memory bank without independent evidence. The MAE pretrained encoder is standard practice and is not counted as a free parameter.

free parameters (6)
  • mask ratio r = 0.4
    Ablated in Tab. 7; 0.3, 0.6, and 0.75 all perform worse. Chosen by hand from a small sweep.
  • inference passes I = 15
    Ablated in Tab. 7; 5 and 10 runs are worse, 20 gives no gain. Selected as the knee of the curve.
  • spatial weight w = 0.3
    Hybrid similarity in Eq. 3; no sensitivity analysis is provided.
  • top-k patches k = 10
    Ablated in supplementary; k=5 loses 0.7 image-AUROC, k=20 gives 0.2 gain at 1.2x cost.
  • Gaussian sigma = 1.4
    Used in Eq. 5 for LAB error computation; no sensitivity analysis is provided.
  • consistency variance threshold = P50 of variance
    50th percentile threshold on variance; the paper reports a 2.4% gain from variance filtering but does not sweep the percentile.
assumptions (4)
  • domain assumption The MAD-SIM renderings and PIAD synt+real images are representative of real pose-agnostic anomaly detection conditions.
    The main benchmark is synthetic or semi-synthetic; the only real dataset (MAD-Real) is relegated to the supplementary and shows degraded performance.
  • domain assumption A 6-parameter affine transform per reference image is sufficient to align reference and query patch grids for cross-view reconstruction.
    STN predicts an affine matrix in Sec. 3.2; the worst-case reference ablation shows degradation with opposite views, indicating the limit of this assumption.
  • domain assumption A pretrained MAE encoder provides features that generalize to object categories in MAD-SIM, PIAD, MVTec, and VisA.
    The encoder is initialized from MAE; no analysis of the pretraining distribution is provided.
  • domain assumption Training on anomaly-free data alone teaches the model to reconstruct anomalous regions as normal content.
    This is the core premise of the method, supported only by empirical results, not by a theoretical argument.
invented entities (1)
  • Anomaly-agnostic tokens (32 global + M local)
    purpose: Memory bank providing anomaly-free reconstruction priors independent of specific anomaly types.
    These are learned parameters without a distinct falsifiable prediction outside the reconstruction output. Their contribution is measured only through the ablation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PADFormer: Pose-agnostic Anomaly Detection from Sparse View Images." pith.science (2026). https://pith.science/paper/OCESDKSU

@misc{pith2026260804210,
  author       = {Pith},
  title        = {Pith review of: PADFormer: Pose-agnostic Anomaly Detection from Sparse View Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OCESDKSU}},
  note         = {Machine review of arXiv:2608.04210}
}
read the original abstract

Pose-agnostic Anomaly Detection (PAD) remains challenging as anomalies can appear under arbitrary viewpoints, requiring methods to handle significant pose variations. Existing approaches rely on complex 3D reconstruction, which are computationally expensive and require extensive multi-view data. We propose PADFormer, a novel image-space approach that leverages Vision Transformer (ViT) to directly reconstruct anomaly-free versions of query images while preserving pose information. Our key insight is to adapt cross-view masked reconstruction for anomaly detection through training exclusively on normal data, combined with dynamic patch selection and spatial alignment mechanisms that enable effective learning from sparse reference views under significant pose variations. During inference, we perform multiple forward passes with different masking patterns to generate an ensemble of anomaly-free reconstructions, ensuring comprehensive coverage of the query image. Anomalies are detected by comparing these reconstructions with the query image. PADFormer achieves state-of-the-art results on the PAD benchmark while maintaining comparable performance on classic few-shot anomaly detection (FSAD) tasks, demonstrating superior efficiency and generalization without requiring 3D reconstruction.

Figures

Figures reproduced from arXiv: 2608.04210 by the authors.

Figure 1
Figure 1. Method overview – We propose PADFormer, a novel ViT￾based method that directly recon￾structs anomaly-free query images from sparse-view references for pose￾agnostic anomaly detection (PAD). PADFormer performs multiple stochas￾tic inference passes to generate diverse reconstructions and error maps, which are averaged to produce final anomaly detection results. In this paper, we introduce PADFormer, a novel approach t… view at source ↗
Figure 2
Figure 2. PADFormer model architecture – PADFormer first patchifies the reference images {Ri|i = 1, . . . , N} into tokens. The query image to be reconstructed Q is ran￾domly masked and also tokenized. Both are processed through two weight-sharing ViT encoders to obtain query feature token fq and concatenated reference features token fref. A Spatial Transformation Network (STN) takes their concatenation to compute transformat… view at source ↗
Figure 3
Figure 3. Within-view pre￾training. An augmented version of the query image is used as reference. Training. PADFormer aims to learn to re￾construct anomaly-free query images based on few-shot reference images. We train our model in two stages to facilitate this objec￾tive. During training, both reference images {Ri |i = 1, . . . , N} and query image Q are anomaly-free. – Within-view training: For each query image Q, we genera… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: PADFormer successfully detects both fine-grained defects (e.g., stains and [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 4
Figure 4. Figure 4: Qualitative results on PAD benchmarks – On the left we visualize the results on MAD-SIM, with missing parts (row 1, 2), burrs (row 3) and stains (row 4). On the right we visualize results on PIAD. See supplementary materials for more qualitative comparisons. 4.4 Ablati…
Figure 5
Figure 5. Figure 5: Qualitative results on FSAD benchmarks: MVTec-AD (L) and VisA (R). before computing differences, which reduces sensitivity to high-frequency noise and edge artifacts. The variance-based consistency filtering (retaining top 50% variance regions) contributes an additiona…
Figure 6
Figure 6. Figure 6: Qualitative results - 15 inference runs and their average error map on MAD￾Sim with 4-shot setup. Tab. 7 analyzes two critical hyperparameters: masking ratio and number of inference runs. The masking ratio of 0.4 achieves optimal performance, balanc￾ing reconstruction …
Figure 7
Figure 7. Figure 7: Qualitative results - MAD-Sim 4-shot setup. Our method can detect anomaly including missing, stains, and burrs effectively [PITH_FULL_IMAGE:figures/full_fig_p026_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results - PIAD 4-shot setup. Our method can generalize on different lighting conditions [PITH_FULL_IMAGE:figures/full_fig_p027_8.png]
Figure 9
Figure 9. Figure 9: Qualitative results - MVTecAD 4-shot setup. On FSAD task, our method can perform fine-grained anomaly detection well [PITH_FULL_IMAGE:figures/full_fig_p028_9.png]
Figure 10
Figure 10. Figure 10: Qualitative results - VisA 4-shot setup. On some challenging cases (row 4), our method may include some noisy predictions due to the reconstruction errors. While it remains effective on very detailed anomalies (row 5, 6) [PITH_FULL_IMAGE:figures/full_fig_p029_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 40 canonical work pages

  1. [1]

    In: International Conference on Learning Representations (2022),https: //openreview.net/forum?id=p-BhZSz59o4

    Bao, H., Dong, L., Piao, S., Wei, F.: BEiT: BERT pre-training of image trans- formers. In: International Conference on Learning Representations (2022),https: //openreview.net/forum?id=p-BhZSz59o4

    work page 2022

  2. [2]

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

    Beizaee, F., Lodygensky, G.A., Desrosiers, C., Dolz, J.: Correcting deviations from normality: A reformulated diffusion model for multi-class unsupervised anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Con- ference. pp. 19088–19097 (2025)

    work page 2025

  3. [3]

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

    Bergmann, P., Fauser, M., Sattlegger, D., Steger, C.: Mvtec ad–a comprehen- sive real-world dataset for unsupervised anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9592–9600 (2019)

    2019

  4. [4]

    In: Proceedings of the 17th In- ternational Joint Conference on Computer Vision, Imaging and Computer Graph- ics Theory and Applications

    Bergmann, P., Jin, X., Sattlegger, D., Steger, C.: The mvtec 3d-ad dataset for un- supervised 3d anomaly detection and localization. In: Proceedings of the 17th In- ternational Joint Conference on Computer Vision, Imaging and Computer Graph- ics Theory and Applications. SCITEPRESS-Science and Technology Publications (2022)

    work page 2022

  5. [5]

    In: Euro- pean Conference on Computer Vision

    Cao, Y., Zhang, J., Frittoli, L., Cheng, Y., Shen, W., Boracchi, G.: Adaclip: Adapt- ing clip with hybrid learnable prompts for zero-shot anomaly detection. In: Euro- pean Conference on Computer Vision. pp. 55–72. Springer (2024)

    work page 2024

  6. [6]

    In: International conference on pattern recognition

    Defard,T.,Setkov,A.,Loesch,A.,Audigier,R.:Padim:apatchdistributionmodel- ing framework for anomaly detection and localization. In: International conference on pattern recognition. pp. 475–489. Springer (2021)

    2021

  7. [7]

    In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers)

    Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidi- rectional transformers for language understanding. In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers). pp. 4171–4186 (2019)

    2019

  8. [8]

    ICLR (2021)

    Dosovitskiy, A., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)

    work page 2021

Show all 59 references
  1. [9]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Fang, Z., Wang, X., Li, H., Liu, J., Hu, Q., Xiao, J.: Fastrecon: Few-shot in- dustrial anomaly detection via fast feature reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17481–17490 (2023)

    2023

  2. [10]

    Advances in neural information processing systems31(2018)

    Golan, I., El-Yaniv, R.: Deep anomaly detection using geometric transformations. Advances in neural information processing systems31(2018)

    2018

  3. [11]

    In: Proceedings of the AAAI conference on artificial intelligence

    Gu, Z., Zhu, B., Zhu, G., Chen, Y., Tang, M., Wang, J.: Anomalygpt: Detecting industrial anomalies using large vision-language models. In: Proceedings of the AAAI conference on artificial intelligence. vol. 38, pp. 1932–1940 (2024) 16 R.Wang et al

    2024

  4. [12]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Gu, Z., Zhu, B., Zhu, G., Chen, Y., Tang, M., Wang, J.: Univad: A training- free unified model for few-shot visual anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 15194–15203 (2025)

    2025

  5. [13]

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

    Gudovskiy,D.,Ishizaka,S.,Kozuka,K.:Cflow-ad:Real-timeunsupervisedanomaly detection with localization via conditional normalizing flows. In: Proceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 98–107 (2022)

    2022

  6. [14]

    In: Ad- vances in Neural Information Processing Systems (NeurIPS)

    Gupta, A., Wu, J., Deng, J., Fei-Fei, L.: Siamese masked autoencoders. In: Ad- vances in Neural Information Processing Systems (NeurIPS). vol. 36 (2023)

    2023

  7. [15]

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

    He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalable vision learners. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16000–16009 (2022)

    2022

  8. [16]

    In: International Conference on Learning Representations

    Hong, Y., Zhang, K., Gu, J., Bi, S., Zhou, Y., Liu, D., Liu, F., Sunkavalli, K., Bui, T., Tan, H.: Lrm: Large reconstruction model for single image to 3d. In: International Conference on Learning Representations. vol. 2024, pp. 50678–50702 (2024)

    2024

  9. [17]

    IEEE Transactions on Neural Networks and Learning Systems (2024)

    Huang, C., Guan, H., Jiang, A., Zhang, Y., Spratling, M., Wang, X., Wang, Y.: Few-shot anomaly detection via category-agnostic registration learning. IEEE Transactions on Neural Networks and Learning Systems (2024)

    2024

  10. [18]

    In: European conference on computer vision

    Huang, C., Guan, H., Jiang, A., Zhang, Y., Spratling, M., Wang, Y.F.: Registration based few-shot anomaly detection. In: European conference on computer vision. pp. 303–319. Springer (2022)

    2022

  11. [19]

    In: European Conference on Computer Vision

    Irshad, M.Z., Zakharov, S., Guizilini, V., Gaidon, A., Kira, Z., Ambrus, R.: Nerf- mae: Masked autoencoders for self-supervised 3d representation learning for neu- ral radiance fields. In: European Conference on Computer Vision. pp. 434–453. Springer (2024)

    2024

  12. [20]

    Advances in neural information processing systems28(2015)

    Jaderberg, M., Simonyan, K., Zisserman, A., et al.: Spatial transformer networks. Advances in neural information processing systems28(2015)

    2015

  13. [21]

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

    Jeong, J., Zou, Y., Kim, T., Zhang, D., Ravichandran, A., Dabeer, O.: Winclip: Zero-/few-shot anomaly classification and segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 19606– 19616 (2023)

    2023

  14. [22]

    In: Proceedings of the 32nd ACM International Conference on Multimedia

    Jiang, B., Xie, Y., Li, J., Li, N., Chen, B., Xia, S.T.: Igspad: Inverting 3d gaussian splatting for pose-agnostic anomaly detection. In: Proceedings of the 32nd ACM International Conference on Multimedia. pp. 10229–10237 (2024)

    2024

  15. [23]

    In: The Thirteenth International Conference on Learning Representations (2025),https: //openreview.net/forum?id=QQBPWtvtcn

    Jin, H., Jiang, H., Tan, H., Zhang, K., Bi, S., Zhang, T., Luan, F., Snavely, N., Xu, Z.: Lvsm: A large view synthesis model with minimal 3d inductive bias. In: The Thirteenth International Conference on Learning Representations (2025),https: //openreview.net/forum?id=QQBPWtvtcn

    2025

  16. [24]

    Scientific Reports15(1), 33929 (2025)

    Jin, J., Xu, Y., He, H., Gao, F., Zeng, W., Wang, W., Guo, B., Xu, Z.: A swin transformer-based hybrid reconstruction discriminative network for image anomaly detection. Scientific Reports15(1), 33929 (2025)

    2025

  17. [25]

    In: European conference on computer vision

    Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and super-resolution. In: European conference on computer vision. pp. 694–711. Springer (2016)

    2016

  18. [26]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Kruse, M., Rosenhahn, B.: Multi-flow: Multi-view-enriched normalizing flows for industrial anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 3933–3944 (2025)

    2025

  19. [27]

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

    Kruse, M., Rudolph, M., Woiwode, D., Rosenhahn, B.: Splatpose & detect: Pose- agnostic 3d anomaly detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3950–3960 (2024) PADFormer 17

    2024

  20. [28]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Li, X., Tan, X., Chen, Z., Zhang, Z., Zhang, R., Guo, R., Jiang, G., Chen, Y., Qu, Y., Ma, L., et al.: One-for-more: Continual diffusion model for anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 4766–4775 (2025)

    2025

  21. [29]

    In: Proceedings of the IEEE/CVF Winter Conference on Ap- plications of Computer Vision

    Li, Y., Goodge, A., Liu, F., Foo, C.S.: Promptad: Zero-shot anomaly detection using text prompts. In: Proceedings of the IEEE/CVF Winter Conference on Ap- plications of Computer Vision. pp. 1093–1102 (2024)

    2024

  22. [30]

    In: European Conference on Computer Vision

    Liu, Y., Hu, Y.S., Chen, Y., Zelek, J.: Splatpose+: Real-time image-based pose- agnostic 3d anomaly detection. In: European Conference on Computer Vision. pp. 378–391. Springer (2024)

    2024

  23. [31]

    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

  24. [32]

    arXiv preprint arXiv:2506.03660 (2025)

    Luo, W., Yao, H., Cao, Y., Chen, Q., Gao, A., Shen, W., Yu, W.: Inp-former++: Advancing universal anomaly detection via intrinsic normal prototypes and resid- ual learning. arXiv preprint arXiv:2506.03660 (2025)

    2025 arXiv

  25. [33]

    In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference

    Ma,W.,Zhang,X.,Yao,Q.,Tang,F.,Wu,C.,Li,Y.,Yan,R.,Jiang,Z.,Zhou,S.K.: Aa-clip: Enhancing zero-shot anomaly detection via anomaly-aware clip. In: Pro- ceedings of the Computer Vision and Pattern Recognition Conference. pp. 4744– 4754 (2025)

    2025

  26. [34]

    In: Proceedings of the Winter Conference on Applications of Computer Vision

    Maack, R., Thun, L., Liang, T., Tercan, H., Meisen, T.: Pcad: A real-world dataset for 6d pose industrial anomaly detection. In: Proceedings of the Winter Conference on Applications of Computer Vision. pp. 1132–1141 (2025)

    2025

  27. [35]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Nafez, M., Koochakian, A., Maleki, A., Habibi, J., Rohban, M.H.: Patchguard: Adversarially robust anomaly detection and localization through vision transform- ers and pseudo anomalies. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 20383–20394 (2025)

    2025

  28. [36]

    In: European Conference on Computer Vision

    Qu, Z., Tao, X., Prasad, M., Shen, F., Zhang, Z., Gong, X., Ding, G.: Vcp-clip: A visual context prompting model for zero-shot anomaly segmentation. In: European Conference on Computer Vision. pp. 301–317. Springer (2024)

    2024

  29. [37]

    arXiv preprint arXiv:2501.03229 (2025)

    Rajasegaran,J.,Chen,X.,Li,R.,Feichtenhofer,C.,Malik,J.,Ginosar,S.:Gaussian masked autoencoders. arXiv preprint arXiv:2501.03229 (2025)

    2025 arXiv

  30. [38]

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

    Reiss, T., Cohen, N., Bergman, L., Hoshen, Y.: Panda: Adapting pretrained fea- tures for anomaly detection and segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2806–2814 (2021)

    2021

  31. [39]

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

    Roth, K., Pemula, L., Zepeda, J., Schölkopf, B., Brox, T., Gehler, P.: Towards total recall in industrial anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14318–14328 (2022)

    2022

  32. [40]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Sadikaj,Y.,Zhou,H.,Halilaj,L.,Schmid,S.,Staab,S.,Plant,C.:Multiads:Defect- aware supervision for multi-type anomaly detection and segmentation in zero-shot learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 22978–22988 (2025)

    2025

  33. [41]

    Computer Vision and Image Understanding241, 103959 (2024)

    Schwartz, E., Arbelle, A., Karlinsky, L., Harary, S., Scheidegger, F., Doveh, S., Giryes, R.: Maeday: Mae for few- and zero-shot anomaly detection. Computer Vision and Image Understanding241, 103959 (2024)

    2024

  34. [42]

    Color Research & Application: Endorsed by Inter-Society Color Council, The Colour Group (Great Britain), Canadian Society for Color, Color Science Association of 18 R.Wang et al

    Sharma, G., Wu, W., Dalal, E.N.: The ciede2000 color-difference formula: Imple- mentation notes, supplementary test data, and mathematical observations. Color Research & Application: Endorsed by Inter-Society Color Council, The Colour Group (Great Britain), Canadian Society fo...

    2005

  35. [43]

    In: International Conference on Learning Representations (2021)

    Sohn, K., Li, C.L., Yoon, J., Jin, M., Pfister, T.: Learning and evaluating represen- tations for deep one-class classification. In: International Conference on Learning Representations (2021)

    2021

  36. [44]

    NeurIPS (2017)

    Vaswani, A., et al.: Attention is all you need. NeurIPS (2017)

    2017

  37. [45]

    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

  38. [46]

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

    Weinzaepfel, P., Leroy, V., Lucas, T., Brégier, R., Cabon, Y., Arora, V., Antsfeld, L., Chidlovskii, B., Csurka, G., Revaud, J.: Croco: Self-supervised pre-training for 3d vision tasks by cross-view completion. In: Advances in Neural Information Processing Systems (NeurIPS) (2022)

    2022

  39. [47]

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

    Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: Simmim: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 9653–9663 (2022)

    2022

  40. [48]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Xu, J., Lo, S.Y., Safaei, B., Patel, V.M., Dwivedi, I.: Towards zero-shot anomaly detection and reasoning with multimodal large language models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 20370–20382 (2025)

    2025

  41. [49]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Yang, K., Cao, J., Bai, Z., Su, Z., Tagliasacchi, A.: Piad: Pose and illumination agnostic anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 4734–4743 (2025)

    2025

  42. [50]

    Sensors24(8), 2440 (2024)

    Yang,Q.,Guo,R.:Anunsupervisedmethodforindustrialimageanomalydetection with vision transformer-based autoencoder. Sensors24(8), 2440 (2024)

    2024

  43. [51]

    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

  44. [52]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Zavrtanik, V., Kristan, M., Skočaj, D.: Draem–a discriminatively trained re- construction embedding for surface anomaly detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 8330–8339 (2021)

    2021

  45. [53]

    Pattern Recognition112, 107706 (2021).https://doi.org/ 10.1016/j.patcog.2020.107706,https://www.sciencedirect.com/science/ article/pii/S0031320320305094

    Zavrtanik, V., Kristan, M., Skočaj, D.: Reconstruction by inpainting for visual anomaly detection. Pattern Recognition112, 107706 (2021).https://doi.org/ 10.1016/j.patcog.2020.107706,https://www.sciencedirect.com/science/ article/pii/S0031320320305094

    2021

  46. [54]

    In: International Conference on Learning Rep- resentations (2022)

    Zhou, J., Wei, C., Wang, H., Shen, W., Xie, C., Yuille, A., Kong, T.: Image bert pre-training with online tokenizer. In: International Conference on Learning Rep- resentations (2022)

    2022

  47. [55]

    Advances in Neural Information Processing Systems36, 44558–44571 (2023)

    Zhou, Q., Li, W., Jiang, L., Wang, G., Zhou, G., Zhang, S., Zhao, H.: Pad: A dataset and benchmark for pose-agnostic anomaly detection. Advances in Neural Information Processing Systems36, 44558–44571 (2023)

    2023

  48. [56]

    In: The Twelfth International Conference on Learning Representations (2023)

    Zhou, Q., Pang, G., Tian, Y., He, S., Chen, J.: Anomalyclip: Object-agnostic prompt learning for zero-shot anomaly detection. In: The Twelfth International Conference on Learning Representations (2023)

    2023

  49. [57]

    In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition

    Zhu, J., Pang, G.: Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In: Proceedings of the IEEE/CVF confer- ence on computer vision and pattern recognition. pp. 17826–17836 (2024) PADFormer 19

    2024

  50. [58]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Zhu, W., Wang, L., Zhou, Z., Wang, C., Pan, Y., Zhang, R., Chen, Z., Cheng, L., Gao, B.B., Zhang, J., et al.: Real-iad d3: A real-world 2d/pseudo-3d/3d dataset for industrial anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 1521...

    2025

  51. [59]

    In: European conference on computer vision

    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. pp. 392–408. Springer (2022) PADFormer 1 Supplementary Material The supplementary docume...

    2022

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.